Growth Hacking 2026: the complete guide

In brief

• The three standout growth hacks of 2025–2026 are product loops, not campaigns: ChatGPT's “Ghibli” wave (1 million sign-ups in one hour according to Sam Altman), the “vibe coding” growth of Lovable and Cursor (Lovable reached $100 million in annual recurring revenue (ARR) about eight months after launch, and Cursor went from $100 million in January 2025 to $500 million in June 2025, with almost no advertising) and the open-source explosion of OpenClaw (more than 100,000 GitHub stars and 2 million visitors in one week according to its launch post of 29 January 2026, followed by a security crisis).
• The search channel is turning into a citation channel: with AI Overviews (the AI-generated summaries at the top of Google results), clicks on an organic result fall from 15% to 8% of visits (Pew Research) and the click-through rate (CTR) of position 1 drops by 58% (Ahrefs, February 2026). So you need to target the short queries that AI assistants themselves type into Google (the “fan-out”), be recommended by third parties rather than by yourself (GEO, generative engine optimisation), and remain usable by agents (MCP, Agent Skills, WebMCP, agentic commerce protocols).
• My position: in 2026, a growth hacker must (1) expose their product to agents, (2) produce citable proprietary data, (3) reinvest in underused physical and human channels (events, handwritten mail, local radio), and (4) draw clear red lines on stolen data, manipulation and undisclosed AI ambassadors, now governed by Article 50 of the European regulation on AI (AI Act), which has applied since 2 August 2026.

Each section opens with a “quick take”, the lesson to remember in a few lines, and closes on what awaits in the next one. From section 8 onwards, the levers are applied to LEEZ, this guide's worked example.

↑ Contents

Section 1

Introduction to growth hacking

Where does the term “growth hacking” come from?

The term was popularised in 2010 by Sean Ellis, who had led Dropbox's early growth, to describe someone whose “true north is growth”. Andrew Chen then spread it in 2012 with the essay “Growth Hacker is the new VP Marketing”. The founding idea: in a start-up with no budget, growth comes first not from campaigns but from the intersection of product, data and marketing.

A working definition

Growth hacking is a method of rapid, measured experimentation, applied to the whole customer journey, to find the most effective growth levers at the lowest cost. It differs from traditional marketing on three points:

• Scope: traditional marketing often stops at acquisition and awareness; growth also covers activation, retention, referral and revenue, and therefore the product itself.
• Method: hypothesis, test, measurement, iteration, in cycles of a few days rather than annual plans.
• Profile: a mix of technical skills (code, data, automation) and an understanding of human behaviour.

In 2026, I expect a revival of growth hacking. Thanks to AI, digital creation is more affordable than ever: building a product is no longer the obstacle. Acquisition, however, remains a challenge, and that is where the method becomes valuable again.

The AARRR framework (“pirate metrics”)

Formalised by Dave McClure in 2007, AARRR splits the journey into five stages: Acquisition (how people find you), Activation (the first successful experience, the “aha moment”), Retention (they come back), Referral (they recommend you), Revenue (they pay). In 2026, I add two implicit layers: discoverability by machines (being cited by an AI or called by an agent) before acquisition, and trust as a condition that cuts across every stage.

How it evolved up to 2026

• 2010–2015: the age of spectacular hacks, studied as textbook cases: Dropbox's referral programme, Airbnb and Craigslist, and before them Hotmail's “PS: I love you” signature (1996).
• 2015–2020: professionalisation, “growth” teams, product-led growth, A/B testing on an industrial scale.
• 2020–2023: saturation of paid channels, rising acquisition costs, the announced end of third-party cookies, tightening of the European GDPR and of the Swiss Federal Act on Data Protection (FADP).
• 2024–2026: generative AI brings down the cost of experimentation and content production, but also brings down the value of generic content; search engines become answer engines; agents become “users” in their own right.

Why the term is contested

The word “hack” has aged badly: it suggests shortcuts, and sometimes borderline practices (unrestrained scraping, spam, dark patterns). Many companies now talk about “growth marketing”, “growth engineering” or “lifecycle”. I keep the term for the reason given at the start: it describes a mindset, finding the asymmetric lever that others have not seen. On one condition: that you own it with explicit ethics. This guide is built on that principle.

Three recent hits show this mindset at work. The first began with simple cartoon-style portraits.

↑ Contents

Section 2

The top 3 growth hacks of 2025–2026

2.1 ChatGPT and the “Ghibli” wave: 1 million users in one hour

Context. In late March 2025, OpenAI built GPT-4o's native image generation into ChatGPT. Within days, social networks filled up with portraits in the Studio Ghibli style.

The mechanics. A feature that turns personal data (your photo) into a shareable, recognisable and flattering object. Every image posted is an advert for the tool, with an instantly identifiable visual signature. The marginal cost of acquisition is close to zero: the user is the medium.

The numbers. On 31 March 2025, Sam Altman wrote on X that when ChatGPT launched, 26 months earlier, the tool had gained one million users in five days, and that it had just gained one million in one hour. Reuters, citing Similarweb, reported that average weekly active users had passed the 150 million mark for the first time that year and, according to Sensor Tower, that downloads of the app had risen by 11% in a week. OpenAI was then claiming around 500 million weekly users. Altman had to ask users publicly to calm down because the GPUs were “melting”.

Why it worked. Three ingredients: a result you can show straight away (visual identity), a shared aesthetic that creates a trend effect, and zero friction (free, in a tool already installed). It is the perfect viral loop: using = sharing.

The dark side. An aesthetic borrowed from a studio whose founder is known for his hostility to generated art opened a copyright debate, which the European decisions of late 2025 (see section 20) make even more sensitive.

What you can replicate. Design a feature whose output is a shareable, identifiable artefact (a card, a score, a portrait, a year in review). Measure the “share rate” per activated user, not impressions. And check the legality of the style or content your feature imitates.

2.2 Lovable and Cursor: “vibe coding” growth without advertising

Context. “Vibe coding” (describing in natural language the application you want and letting an AI write it) is the fastest-growing software category in recent history.

The numbers.

• Lovable (Stockholm) raised $200 million in July 2025 at a valuation of $1.8 billion, reached about $100 million in ARR about eight months after launch, then $200 million in ARR in November 2025 according to Bloomberg, before a $330 million Series B at a valuation of $6.6 billion in December 2025 (TechCrunch). The company said it was approaching 8 million users in November 2025.
• Cursor (Anysphere) went from $100 million in ARR in January 2025 to $500 million in June, $1 billion in November 2025, then more than $2 billion in early 2026 according to Bloomberg and TechCrunch.

The mechanics. A product loop in which every user creates a public artefact (an application, a site, a repository) that demonstrates the tool's value. According to Bloomberg, quoted by Entrepreneur (April 2025), in January 2025 Cursor became the fastest product to reach $100 million in ARR without spending a cent on marketing, with almost all its revenue then coming from individuals, driven by developer word of mouth on X, Reddit, Hacker News and YouTube. Credit-based freemium creates a natural conversion: you pay when you are already building.

Why it worked. The “time-to-wow” is measured in minutes; the demonstration happens in public; the community of creators becomes the acquisition channel.

The limits. According to Barclays analysts quoted by TechCrunch (November 2025), Lovable's traffic had fallen by 40% in September 2025 and that of v0 (Vercel) by 64% since May, a reminder that retention remains the real judge. Valuations above 30 times ARR are bets, not facts.

What you can replicate. Cut the time to first success to under five minutes; turn everything a user creates into a public showcase (badge, gallery, shareable link); measure retention by cohort before celebrating acquisition.

2.3 OpenClaw and Moltbook: open-source virality and its downside

Context. OpenClaw is an open-source personal AI agent, hosted locally and controlled through messaging apps (WhatsApp, Telegram, Signal, Discord), created by Peter Steinberger in late 2025. It changed its name twice in a few days (Clawdbot, then Moltbot on 27 January 2026 after a trade mark infringement complaint from Anthropic, then OpenClaw a few days later).

The numbers. According to OpenClaw's launch post of 29 January 2026, the repository had already passed 100,000 GitHub stars and drawn 2 million visitors in a single week; it had 247,000 stars and 47,700 forks as of 2 March 2026. Moltbook, a social network reserved for agents and launched on 30 January 2026, claimed 1.4 million active agents in early February. On 15 February 2026, Steinberger joined OpenAI and the project moved under an independent foundation.

The mechanics. A rare combination: a tangible promise (“an assistant that acts for you through WhatsApp”), a story told in episodes (the name changes, the trade mark dispute) and a public spectacle (Moltbook, where humans can only watch agents talk to each other). Every twist revived attention.

The downside. Within a few weeks: a remote code execution vulnerability (CVE-2026-25253), tens of thousands of instances exposed on the internet (Censys counted more than 21,000 at the end of January) and malicious extensions in its “skills” marketplace. In March 2026, the Chinese authorities restricted its use in state-owned enterprises.

Why it is a textbook case. It shows the power of a serialised story and a developer community, but also that growth without security destroys the trust of businesses within days.

What you can replicate. Serialise your launch; show the product in action (a public “show”); and budget for security before the peak, not after.

These three hits look like happy accidents. Yet you can bring about more modest ones methodically, and the method comes down to a loop in three steps.

↑ Contents

Section 3

The mechanics in one chart: ideation, experimentation, analysis

The loop reads in the direction of the arrows: each analysis produces a lesson that becomes the next hypothesis, and the North Star metric, the single indicator that sums up the value delivered to customers, decides what deserves to be tested.

The disciplines to know

My advice: nobody masters everything. A good growth hacker is an expert in one area, solid in a second, and able to talk to the third. In 2026, AI agents fill some of the technical gaps, which makes judgement and ideation more decisive.

One point that changes the mindset: most experiments fail. At Microsoft, about a third of the ideas tested improve the target metric; at Bing, 15%; at Booking.com, 10%, according to figures compiled by Ronny Kohavi. A high failure rate is normal, and a success rate that looks too good often signals tests that are too cautious, or false positives.

A loop is judged in the field. My own fieldwork began with an auction platform that was losing its audience, and a burger passport.

↑ Contents

Section 4

Case study: Cronodeal SA and Burgerpass

The problem: neither customers nor products

Cronodeal was a Swiss platform for descending-price auctions, created in late 2013 by Alessandro Soldati in Bussigny (Vaud) and inspired by the Aalsmeer flower auction in the Netherlands: the price of a product fell over time, until a buyer made up their mind. I worked there from 2015 to 2017 as a partner, in charge of growing the marketplace.

When I arrived, we had the typical marketplace problem: not enough customers to get good products, and not enough good products to attract customers. According to PME magazine, the platform had seen peaks of 20,000 unique users a day at launch; four months later, only 500 very loyal visitors remained. The daily 24 heures, for its part, reported 17,000 users won over.

The method: ideation, analysis, MVP

I had just come out of the HES-SO marketing and innovation programme, set up by Nathalie Nyffeler. I applied its product creation steps: ideation, analysis, then MVP (minimum viable product). The idea we chose rode the gourmet burger wave of the time, started by Holy Cow in Lausanne in 2009.

The product: a passport for gourmet burgers

The Burgerpass principle fitted in one line: at each partner restaurant, two burgers for the price of one, with the rest of the bill paid by the customer.

It was also a simple way to promote local produce. We only accepted restaurants that served gourmet burgers, meaning those that bought from the butcher, the market gardener, the cheesemaker and the baker.

The launch: a price set by auction

Six months later, I held the first Burgerpass in my hands. Two weeks after that, we put it on sale on Cronodeal, so that the descending-price auction would set its price: 1,100 copies sold in 24 hours, at 58 Swiss francs on average. It was the start of a revival for Cronodeal, and of five annual editions for the Burgerpass.

The press documented what came next. According to Largeur, the Burgerpass was designed as a growth hacking solution to counter stagnating sales by strengthening the community around the site. A crowdfunding campaign on wemakeit raised a little over 4,600 francs. The community reached about 4,500 people according to the programme of the HES-SO Valais “Business Ideas” event, and about 50 partner establishments for the fourth edition according to PME, which reported revenue of 450,000 francs for Cronodeal in 2016. The project was taken over by Michael Dusong on 13 May 2019. The model was adapted into Escapegamepass (31 affiliated escape games, 327 teams at the Swiss championship), since sold to the Groupement des escape games romands.

What I take from it: the Burgerpass is a case of “community as a hook product”, a physical, local, collectible object that creates recurring traffic to a platform that lacked it. It is exactly the kind of lever that AI does not commoditise (see sections 22 and 33).

A burger almost sells itself. The next case is the opposite: a service nobody wants to hear about, because it deals with death.

↑ Contents

Section 5

Case study: Tooyoo SA

The problem: a subject nobody wants to talk about

Tooyoo is a Swiss digital platform that lets people document their last wishes (living will, organ donation, management of online accounts, estate matters) and pass them on securely to their loved ones after their death. Initiated by La Mobilière (Mobilière Services SA) and developed at the EPFL Innovation Park, it was launched at the end of July 2017 in French, German and English. It won the Gold Award for Innovation at Best of Swiss Web 2018.

I joined La Mobilière's innovation unit in March 2018 to work on Tooyoo, with the title “Digital Happyender in Chief”; co-founder Ralph Rimet held that of “Executive Happyender in Chief”. Our answer was a complete growth engine, which brought in thousands of users.

Paper first

We launched campaigns inviting people to download and print the important documents, in order to make the brand known among older people. The result: more than 100,000 downloads of advance care directives, living wills and wills. The software first became known through paper.

A taboo broken on LinkedIn

The other campaign ran through the LinkedIn profile of Anna Aker, Tooyoo's Chief Ambassador Officer, and her more than 10,000 followers. The end of life is a delicate subject; we addressed it in a professional setting. Some posts reached several million views: they broke a taboo in an environment that did not expect this kind of post.

The results

The platform passed 20,000 members in 2019. The solution is now distributed by Groupe Mutuel throughout Switzerland. The company was bought out by its management in 2023 (MBO): Ralph Rimet is still its CEO and one of its partners, alongside the other managers who ran the project at the time.

What I take from it: “low-frequency, emotionally charged” products grow through trusted third parties (insurers, notaries, the media) and through useful content, far more than through direct advertising. In 2026, it is also an ideal candidate for GEO: questions such as “how to write a living will in Switzerland” are put to AI assistants.

Tooyoo at least had an audience that knew the subject concerned them. At Tayo, nobody was asking for anything.

↑ Contents

Section 6

Case study: Tayo SA

The problem: a sector that asked for nothing

Tayo (Tayo Software SA) is a start-up from French-speaking Switzerland, founded by Etienne Friedli and Claude Frey, that develops a cloud solution for the technical management of property portfolios, connecting property managers, tenants, caretakers and suppliers on a single platform. I was its Head of Growth. Retraites Populaires announced a partnership with Tayo to fully digitalise the technical management of its rental portfolio.

The engine: a complete media outlet

Our answer was to create a complete media outlet: Tayo Tayo, “Le Média Immo”, the growth engine of Tayo SA. It reached more than one million Swiss people on LinkedIn, including thousands of property professionals. The series “Moteur!” and the artwork “Rohr Kreuz 01” came out of it.

The “Moteur!” series

I directed the web series “Moteur!”, shot directly in the offices of the heads of property management firms, our clients. It reached 1,200,000 people on its own.

Another Tayo Tayo production: the artwork “Rohr Kreuz 01”, created with Studio Raphaël Lutz to illustrate Tayo's source code. It was unveiled exclusively at IMMO22. On 22 February 2022, I launched it on Product Hunt, where it ranked 30th worldwide for the day across all categories, and 1st in design (79 votes). A documentary is devoted to it.

What I take from it: in niche B2B, “human” content (a video series with real clients, an art object) creates social proof that AI cannot manufacture credibly. It is a media asset (section 26) as much as a commercial asset.

These three cases belong to the past. To put the levers of 2026 to the test, I needed a present-day testing ground, and the most demanding one possible.

↑ Contents

Section 7

LEEZ: the worked example running through this guide

The goal

Everything LEEZ undertakes serves a single goal: to allow thousands of Swiss entrepreneurs to retire with peace of mind, and to give a new generation the chance to take over their businesses and keep them going.

The stakes go beyond the platform. SMEs make up more than 99% of the country's companies and provide two thirds of its jobs, according to the Federal Statistical Office. When a healthy SME closes for lack of a buyer, jobs, know-how and customers disappear with it. When it is transferred, it keeps all of that, along with the lead it had over its competitors.

The figures show the scale of it. More than one Swiss SME in six is currently looking for a successor, according to the president of the Swiss umbrella organisation for business succession. About a third of successions fail, according to the platform KMU Next, and Zürcher Kantonalbank estimates that a third of SMEs are liquidated when the time comes to settle their succession. Yet a business that is taken over holds up better than a newly created one: more than 95% are still active five years later, against fewer than one in two new businesses, according to the Centre Patronal.

All of this is what LEEZ aspires to improve. Behind these figures are thousands of Swiss people who have devoted whole stretches of their lives to their business. That work deserves respect: the respect of not letting it die out for lack of a buyer.

What LEEZ is

LEEZ (leez.ch) is a Swiss marketplace for business transfers. It connects SME owners who want to sell their company with buyers, and with the experts who support them. It was founded by four Swiss entrepreneurs who had each started, grown and then passed on their own SME, and who had seen how complex the process was, and how little visibility it had.

What it does

In early October 2026, the site shows about 3,300 businesses for sale in 25 cantons, more than 300 registered buyers and four languages.

What it wants to become

LEEZ's stated ambition comes down to three points: give more visibility to businesses looking for a successor, bring the whole ecosystem together on a single platform, and offer sellers, buyers and intermediaries a simple, structured and secure tool. The aim is to become the go-to player for business transfers in Switzerland. LEEZ puts figures on the stakes as follows: more than half of Swiss business leaders are over 55, and about 75,000 businesses, or some 400,000 jobs, will have to change hands by 2030. Other estimates exist: Dun & Bradstreet counted 90,667 SMEs in need of a successor in 2025.

My reading, from the growth side: move from a listings site to the infrastructure of business transfers. The one that AI assistants cite, and whose data sets the standard.

Who lists the most businesses for sale: the platform ranking

To put LEEZ in context, I counted the businesses for sale shown on each buying and selling platform. Here are the ten largest listing inventories I found, recorded on the sites on 5 October 2026. I kept only platforms that publish their own listings, not aggregators that republish other people's.

This ranking comes with four caveats.

• A listing is not an SME. The largest inventories are made up of small retail businesses, cafés and restaurants. On CessionPME, the section for companies for sale has 3,361 listings, against 58,826 for retail businesses.
• The same listings circulate from one site to another. A business broker publishes the same mandate on several platforms. These inventories cannot be added together.
• The counters do not measure the same thing. Some include franchises to be set up, commercial premises or searches for investors. Three figures are my own calculations based on the sites' counters: CessionPME, BizBuySell and Place des Commerces.
• The inventory moves by the hour. BusinessesForSale.com showed 51,879, 51,906 and then 51,949 listings on the same day.

Just below the top 10 come Bsale in Australia (13,595 listings) and Daltons Business in the United Kingdom (13,208). In Japan, BATONZ claims more than 10,000 deals online.

Outside the ranking, aggregators show more: Kumo shows 101,453 and Bpifrance's Bourse de la transmission 42,054. Exchanges focused on SMEs are much smaller: 6,387 offers on nexxt-change in Germany, 2,008 on Fusacq in France, 1,189 on Brookz in the Netherlands.

The Swiss ranking

I did the same count in Switzerland on 5 October 2026, on marketplaces and on business brokers that publish their catalogue online.

Foreign platforms carry little weight in Switzerland: 73 Swiss listings on DealStream, 17 on BusinessesForSale.com.

The caveats of the world ranking apply here too. Three figures were not read straight off a counter: those for Remicom and Nachfolger-Börse are estimates, and the one for Relève PME comes from a presentation.

My reading: LEEZ shows the largest inventory in the country, more than eight times that of the second marketplace. It is a figure nobody else can publish, and therefore a reason to be cited. But an inventory does not say what sells: no Swiss marketplace publishes a figure for completed transfers. The first to do so will hold the figure that the press and AI assistants will repeat.

Why a marketplace is the most demanding case

How to read what follows

From here on, the levers are applied to LEEZ in two steps. First, what the lever brings to the goal: how it helps an owner pass on a business, or a buyer find one. Then what that looks like in practice.

I did not choose this case for convenience. A tactic that works on a marketplace, with two audiences whose interests are opposed and who must come to agree on every point, will work on a simpler model. The reverse is not true.

First lever, and the most powerful: the one that carried the three hacks in section 2. It remains to be seen whether it works in a trade where nobody shares anything.

↑ Contents

Section 8

Product-led growth

What the three hacks have in common

In each of these three cases, using the product already means spreading it.

What “product-led growth” means

The term product-led growth was coined in 2016 by Blake Bartlett, of the fund OpenView. In this model, the product itself does the work of acquisition, activation and retention. Conventional growth hands that work to advertising or a sales force.

Three conditions keep coming up in the cases that work.

• Immediate value. The user gets a first result within minutes, without talking to anyone.
• One use that leads to another. Each use exposes the product to someone new: an image shared, an invitation sent, or a person who has to come onto the platform to respond.
• A clear journey. The user knows where they stand, what comes next and what it costs.

The limits

• Retention remains the judge. A product that spreads fast can empty just as fast: Lovable's traffic showed as much (section 2).
• Trust is lost in a few days. OpenClaw grew faster than its security, and companies turned away from it.
• Not everything can be shared. An owner will never post on social media that they are selling their business. In a confidential trade, the loop does not run through public sharing. It runs through the people the process forces you to bring in: the counterparty, the advisers, the experts.

My advice

Before investing in a channel, I ask the product three questions. What does the user get in their first five minutes? Who else has to come onto the product for them to move forward? Does the user know where they stand, what comes next and what it costs?

Applied to LEEZ

What it brings. Transferring a business is expensive, partly because the market is opaque. The seller knows neither which steps lie ahead, nor which expert they need, nor what that expert should cost. Neither does the buyer. A tool that makes the path clear and puts experts in competition saves sellers and buyers substantial costs, and lets more transfers go all the way through.

The orders of magnitude. For a sale of 15 million Swiss francs, expect about 300,000 francs in transfer costs, or 2% of the price. For the smallest transfers, it is generally 10% of the final sale price, with a minimum of 15,000 francs, commission and expenses included. In proportion, the small business therefore pays five times more than the large one.

These amounts vary widely with the sector, the complexity of the due diligence (the in-depth audit of the business before the purchase), the number of employees and the preparation done beforehand by the seller and the buyer. Handing your business over to an employee or to your managers lightens both the bill and the support needed, because much of the knowledge is already in the successors' hands.

What opacity costs. LEEZ's 2026 reference guide to acquisition financing reviewed 141 accounts from sellers, buyers, banks, business brokers and tax advisers. Four findings emerge.

• Prices are not published. None of the named profiles of banks, guarantors and business brokers consulted shows a price. The firm Auditrium is one of the few to put a figure on an adviser's success fee: often 2 to 5% of the price for a mid-sized business.
• Timelines are long. Selling to a third party takes five to nineteen months in the cases reviewed. A family succession lasts more than six years on average, according to Zürcher Kantonalbank.
• Mistakes are costly. In a case published by VZ, a seller who started too late could no longer buy in pension fund years and missed out on nearly 100,000 francs in tax savings. In a case decided by the Federal Supreme Court (judgment 4A_220/2013), a warranty given on the order book cost the seller 1.39 million francs.
• Neither side knows what the other would accept. Two thirds of sellers would be willing to finance part of the price, but only one buyer in eight considers asking for it, according to the KMU Spiegel 2015 study by the St. Gallen University of Applied Sciences.

Why it works on a marketplace. Product-led growth needs volume, and a marketplace has it: about 3,300 businesses listed, more than 300 registered buyers. Volume brings the introductions. What matters next is the experience once the seller and the buyer have been put in touch. That is what creates the “wow” effect, keeps both parties on board and prevents disintermediation.

What is already automated. Part of the journey already happens without paper or appointments. That is useful from the very first step: according to the Observatoire BCV des entreprises 2025, one in two Vaud businesses due to be transferred has not yet been valued.

The non-disclosure agreement shows what automation changes. In a conventional sale, it is the first speed bump: each interested buyer receives a document, signs it by hand and sends it back, and the seller or their business broker keeps the list of who signed what. On LEEZ, the verified buyer signs online. Each access to the file is thus tied to a verified identity and a signed agreement. Confidentiality becomes a step in the product, and no longer a formality between two emails.

Four languages. LEEZ works in French, German, Italian and English. This lets it cover the whole of Switzerland, bring together a seller from French-speaking Switzerland and a buyer from German-speaking Switzerland, and attract foreign capital for larger acquisitions.

In practice. LEEZ's ambition is to build the best digital business transfer tool in Switzerland. Two building blocks are in development.

1. A personalised journey, visible to both sides. The experience adapts to the size of the business, its legal form and its situation. The legal form really does change the path: the second pillar (occupational pension), for example, can finance the takeover of a sole proprietorship, but not the direct purchase of shares in a limited company (SA). The full journey is visible to the seller and the buyer alike. Each knows what they have to do, what the other has to do and when. That brings clarity and certainty, at a time when nobody wants to get it wrong.
2. A call for tenders tool, at every step. Some steps require specific expertise, due diligence for example. The entrepreneur posts what they need, the experts connected to the platform respond, and the entrepreneur chooses among several offers. The call for tenders is anonymous: the identity of the seller or buyer is revealed only when they accept an offer, to preserve maximum discretion.

Supported or solo. Each party chooses. The seller, like the buyer, can be supported by a professional from start to finish, or run the transfer themselves and launch a call for tenders only at the steps where they need an expert.

Under 500,000 francs: the same care. An SME worth 15 million can afford to surround itself with help: a business broker, a lawyer, a tax adviser, and access to selected buyers. A business worth less than 500,000 francs cannot, and it is the one that pays the most in proportion. Yet it is the heart of the market: about three quarters of the listings on LEEZ ask for less than 250,000 francs (section 16). LEEZ wants to develop a product for these small businesses that gives them a quality of service and access to buyers close to what an SME worth 15 million gets.

The loop. Each deal brings its counterparty and its advisers onto the platform. Each call for tenders gives experts a reason to come back, and to bring their own clients with them. The more experts there are, the wider the choice and the clearer the prices, which attracts new sellers and new buyers.

This loop answers two difficulties described in section 7. Disintermediation: the deal stays on LEEZ because its steps take place there. Repeat business: it comes from the experts, who return for every deal.

A clear journey is still not enough to close a deal, though. Between the committed buyer and the signature, one obstacle remains that no listing solves: money. To remove it, you have to get inside the trade.

↑ Contents

Section 9

Financing an acquisition: understanding the needs

Why financing is the sticking point

Banks ask for at least 30% equity, part of which must not come from pension savings. Many buyers do not have it. According to a Basler Kantonalbank study published in 2025, 37% of SMEs in north-western Switzerland have not yet resolved the financing of their succession. No single instrument is enough: an acquisition is always financed by a combination.

The benchmarks in this section come from LEEZ's 2026 reference guide to acquisition financing, which describes 22 methods. I have kept seven. They are orders of magnitude, and no substitute for tax advice. EBITDA, which comes up often, is earnings before interest, taxes, depreciation and amortisation.

Seven ways to raise the funds

How a deal structure is put together

Why go into this much detail in a growth guide? Because you only find a weak point in a system you understand. A growth hacker who knows the sector, the field and the trade in depth sees where the journey really gets stuck, and that is where they act. Without meta-vision, that is, without an overall view of the system, no hack is possible: all that remains is tactics applied at random.

In the vast majority of Swiss SME acquisitions, the standard structure is the same. The seller sells shares held as private assets, which in principle makes the capital gain tax-exempt. The buyer purchases them through an acquisition holding company, financed by about 30% equity, 50 to 70% bank debt and 10 to 30% vendor loan or earn-out. This exemption rests on one condition: for five years, the debt must not be repaid with the reserves not needed for operations that existed at the time of the sale.

The proportions change with the size of the deal.

An example in figures: an SME valued at 3 million francs, or five times its EBITDA of 600,000 francs. The buyer puts in 900,000 francs, the bank lends 1.8 million and the seller grants a loan of 300,000 francs. The bank debt is repaid over six years, at 300,000 francs a year.

The canton changes the bill

The canton does not rule out any method, but it changes what each one costs. In the example above, corporate profit tax falls from 102,700 francs a year if the business is in Bern to 58,300 francs if it is in Lucerne. Over six years, the gap comes to about 266,000 francs, or nearly 15% of the bank debt.

The seller's retirement is part of the structure

The sale price is often the seller's retirement capital. A good structure therefore also looks at what the seller has left after tax, and that needs preparing up to ten years in advance.

• Sell shares rather than a sole proprietorship. The gain on privately held shares is in principle tax-exempt. The conversion into a limited company (SA) must take place at least five years before the sale.
• Buy in pension fund years. These buy-ins are deductible from income, provided the capital is not withdrawn within the following three years.
• Take out what is not needed for operations. Separating out the property and the surplus cash reduces the price the buyer has to finance.
• Distribute profits regularly rather than hoarding them.

The right structure reconciles two needs: what the buyer can repay, and what the seller needs for retirement.

In the field: what I heard at the Foire du Valais

A good growth hacker scours their sector and observes it. They bring back accounts and lessons, then turn them into answers to needs, in the form of a product. On 2 October 2026, I attended the Rendez-vous de la transmission d'entreprise at the Foire du Valais. Here are my notes, and what I draw from them.

It is this kind of listening that feeds a good product, like the ones described in this section and the previous one.

Applied to LEEZ

What it brings. A buyer who has pulled off one deal structure can pull off others. They become a multipreneur: an entrepreneur who has bought several businesses. Each financed acquisition paves the way for another, and each seller gains one more buyer.

In practice. The transfer journey (section 8) is meant to bring these parameters together in one place: the size of the business, its legal form, the cantons involved, the buyer's funds and the seller's retirement needs. The aim is to optimise the whole, not each element separately.

An AI of LEEZ's own. An artificial intelligence trained on the transfers completed on LEEZ will refine the experience as deals close. Each transfer improves the recommendations, which brings more transfers.

None of this happens without volume. That is why the rest of this guide focuses on growth and customer acquisition. And in 2026, all customer acquisition starts with the same question: what has AI really changed?

↑ Contents

Section 10

Growth hacking in the age of AI: the full picture

What AI changes, step by step

Acquisition. Search is becoming conversational: AI Overviews and assistants answer without sending a click (see section 15). Generic content mass-produced by AI loses value; original data, expert opinions and brand mentions gain it. Browsing and shopping agents become a new “segment” to serve.

Activation. Onboarding can be personalised in real time by a built-in assistant; “time-to-value” gets shorter. Lovable and Cursor (section 2) deliver their first result within minutes.

Retention. AI makes it possible to detect churn (customers leaving) earlier and to automate relevant follow-ups; but AI summaries in inboxes reduce the visibility of emails (section 21).

Testing. You can generate dozens of page or ad variants in an hour. The bottleneck becomes the traffic available to reach statistical significance, and the judgement to choose what to test.

Cost of experimentation. A prototype that used to take two weeks of development can be built in a day with Cursor, Claude Code or Lovable. The consequence: speed is no longer a lasting competitive advantage; distribution and trust are.

The go-to tools in 2026

• Building: Cursor, Claude Code, Lovable, v0, Replit.
• Automation: n8n, Make, Zapier, and agents able to call MCP servers.
• Research and content: ChatGPT, Claude, Gemini, Perplexity, with Agent Skills to standardise procedures.
• Measuring AI visibility: a new category of tools that track citations in generative engines (section 15).

The limits (to be taken seriously)

• Saturation: when everyone can produce, average content becomes noise.
• Trust: consumer preference for AI-generated creator content fell from 60% in 2023 to 26% in 2025 (section 30).
• Regulation: since 2 August 2026, Article 50 of the AI Act has required chatbots and synthetic content likely to mislead to be disclosed (section 30).
• Hidden costs: model calls, agent errors and security (the OpenClaw case) come at a price.
• Platforms: Google, Meta and the marketplaces keep adjusting their rules against low-quality content and abusive automation.

My verdict: AI is a multiplier, not a strategy. It multiplies good positioning and bad positioning alike.

Applied to LEEZ

What it brings. An owner thinking of selling first wants a number, without exposing themselves or paying an expert. By making the valuation immediate, AI turns the first step into a five-minute matter. That is often what separates a prepared retirement from a forced closure.

In practice.

• The free valuation is the moment when the seller receives something useful. That is what AI should speed up and explain, not the mass production of listings.
• Listings are translated into four languages, with human review. An SME in French-speaking Switzerland can thus find its buyer in Zurich or Ticino.
• An AI already answers “how much is my business worth?” without going through LEEZ. Its answer needs to rest on a reliable source (section 16).

If AI writes, tests and answers in the website's place, what is a website still for? One thing only, and the 2026 numbers point to it.

↑ Contents

Section 11

How people still reach a website in 2026

Mobile or desktop

Share of page views in September 2026, according to StatCounter.

Three takeaways:

• Switzerland remains a desktop country. It is one of the markets where the desktop screen still leads the phone, and the gap widens in B2B, where decisions are made at work.
• Mobile is mostly apps. Most of the time spent on a phone happens in apps, not in the browser. A mobile site is viewed between two apps, quickly, often from a link someone sent.
• Mobile discovers, desktop closes. Industry benchmarks give a conversion rate two to three times higher on desktop. That is an order of magnitude, to be measured in your own analytics.

Where visits come from

One detail matters for website design: according to SE Ranking, 60% of visits sent by an AI land on the home page, compared with 17% for conventional search. The home page is becoming a landing page again.

In Switzerland

IGEM's Digimonitor 2026, based on 1,957 people surveyed in the spring, gives the local picture.

• 80% of the population uses AI, compared with 40% in 2024.
• 77% use the AI mode of conventional search engines.
• 57% search directly in an AI platform, more than on YouTube (34%) or social media (28%).
• ChatGPT reaches 67% of the population, or 4.3 million people, including 17% every day. Gemini follows at 32%, Copilot at 30%.

What I conclude

1. The website is no longer the first page read, it is the last. People come to it to act. Action pages (valuation, quote, alert, contact) deserve more care than the blog.
2. You have to win twice. Once in the answer from Google or the AI, to be named. A second time on the website, to convert the person who remembered the name.
3. Measurement changes. Organic sessions fall even when brand awareness rises. I track brand searches, direct traffic and AI citations, and I filter out bots.

Applied to LEEZ

What it brings. The buyer is often an employee searching in the evening, on their phone. The seller, meanwhile, prepares a substantial file at their desk. If each finds the platform on the screen they use, no possible match is lost for a technical reason.

In practice. This is a hypothesis to test: alerts and listing pages designed for mobile, a valuation and a listing submission designed for desktop. And since 4.3 million people use ChatGPT in Switzerland, the question “how do I sell my business?” is now put to an AI before it is put to Google. Being the cited answer there is the subject of section 16.

The 2026 visitor therefore arrives late, and already informed. What remains is to know what to show them, and at what moment.

↑ Contents

Section 12

Credibility of a digital platform: what to show, to whom and when

Why it has become decisive

• The visitor arrives to verify. They have already done their research with Google or an AI (section 11). The website must confirm, not seduce.
• What the brand says about itself carries little weight. Edelman's 2026 Trust Barometer concludes that the word of unpaid third parties counts far more than the brand's.
• The absence of third-party proof arouses suspicion. In a 2026 TrustedSite survey, 40% of shoppers say they distrust an unknown website that displays no trust indicator verified by a third party. TrustedSite sells these badges: I take the gist, not the figure.
• Machines reason the same way. AI assistants mostly cite sources the brand neither owns nor pays for (section 29).

What to show, and when

Institutional partner logos

• Written permission for each logo.
• The exact word: “partner”, “member” and “supported by” do not mean the same thing, and the Unfair Competition Act (UCA) penalises misleading statements.
• Mirrored proof: the logo links to the page where the institution mentions you. Without it, the logo proves nothing.
• Local first: for an SME in French-speaking Switzerland, a cantonal chamber of commerce reassures more than an American media outlet.

The downloadable security audit: yes, at two levels

• Freely accessible: an attestation or a summary of the penetration test (date, scope, provider, overall conclusion), the certifications obtained, the list of subprocessors, an address for reporting a vulnerability.
• On request, under a non-disclosure agreement (NDA): the full report. Published as it stands, it serves as a map for an attacker.
• Dated and renewed: a 2023 audit displayed in 2026 does you a disservice.
• Nothing you do not have: no label, and no certification in progress presented as obtained.

In Switzerland, hosting in the country and labels such as swiss made software or the Swiss Digital Trust Label speak to local customers.

The white paper and the roadmap, web3 style, for a web2 solution

Web3 projects publish a white paper, a quarter-by-quarter roadmap and code audits because they ask strangers for money without having a track record. The document makes up for the lack of a past. The method has also produced its excesses: roadmaps never delivered and white papers that were nothing but brochures.

My view: for a platform that handles serious business, the methodology white paper is relevant, because it serves three readers at once: the customer who verifies, the journalist looking for a source, and the AI looking for a text to cite. The delivery log is worth more than the roadmap: it proves you are moving forward, where the roadmap only promises. A public roadmap also informs competitors, and every delay shows.

Applied to LEEZ

What it brings. Selling a business that is a life's work requires total trust. At the slightest doubt, the owner gives up, postpones, and sometimes ends up closing. Each piece of proof placed at the right moment removes one reason to give up.

In practice.

• On the home page: a dated listing counter, and the logos of professional associations as partnerships are signed.
• Before the valuation: a sentence on anonymity, saying what a buyer will and will not see.
• Before submitting a listing: a security page, with the audit attestation, the subprocessors and the company's identity.
• A white paper, “How LEEZ values a business”, so that anyone can verify the method.
• A reference guide, modelled on the 2026 edition of LEEZ's guide to acquisition financing: it cites its legal bases in full, links each of its 141 accounts to its source and devotes a chapter to its own limits.

All this proof is addressed to a human reader. But the next visitor may not read the page: it will scan it in a fraction of a second, because it is a program.

↑ Contents

Section 13

What an AI agent is and how to build websites that speak its language

What is an AI agent?

An AI agent is a system built on a language model that receives a goal, plans steps, uses tools (browser, APIs, files, payments) and acts semi-autonomously until it reaches the result or asks for approval. The difference from a chatbot: the chatbot answers, the agent acts.

In 2026, three families of agents visit your website:

1. Generative engine crawlers (for training or real-time search).
2. Browsing agents built into browsers and assistants, which read and click on a user's behalf.
3. Transactional agents that compare, fill a cart and pay via dedicated protocols.

How an agent “reads” a website

An agent prefers explicit structures to visual interfaces. It reads the HTML, the structured data, the accessibility attributes and, increasingly, dedicated entry points (APIs, MCP servers, tools declared via WebMCP). A website that relies on heavy JavaScript, unlabelled buttons or content held in images is partly unreadable to an agent.

The building blocks for speaking the agents' language

Agentic commerce in 2026

Three protocols are competing for agent purchasing: ACP (OpenAI and Stripe), UCP (Google and Shopify) and AP2 (Google, for payment mandates delegated to agents). Section 14 covers the first two in detail. Standardisation is not finished: betting on a single protocol remains risky.

Good practice in concrete terms

1. Audit your website with a browser that has JavaScript turned off: what you see is roughly what an agent in a hurry sees.
2. Add complete structured data (Organization, Person, Product, Article, FAQPage, BreadcrumbList).
3. Write a clear policy in robots.txt: allow search crawlers (for example OAI-SearchBot, PerplexityBot, Googlebot), and make a conscious decision about training crawlers (GPTBot, ClaudeBot, Google-Extended).
4. Publish a minimal API or MCP server for the key actions (search, book, request a quote).
5. If you do e-commerce, follow UCP and ACP through your platform (Shopify, Stripe) rather than building them yourself.

My view on real adoption: in French-speaking Switzerland, the vast majority of SME websites are not even up to date on schema.org. So that is where the lead is to be gained, well before WebMCP or purchasing protocols.

Applied to LEEZ

What it brings. Tomorrow, a young buyer will ask their assistant: “find me a joinery business to take over in the canton of Fribourg”. If the agent cannot read LEEZ's listings, the joinery stays invisible, and it will close for want of being seen.

In practice.

• Each listing can be read without JavaScript, with its structured data: canton, sector, price range, date of last update.
• The robots.txt file is a business decision. The door is open to AI search crawlers, which cite and send visitors back. It is closed to training crawlers, which hoover up listings and give nothing in return: this is the traffic diversion described in section 7.
• An agent can search for a listing and prepare a request for a non-disclosure agreement (NDA). The signature stays human, because that is what protects the seller.

Reading a website is only the first step. Four protocols now let agents act, each through a different door, and not all of them lead to you.

↑ Contents

Section 14

MCP, skills, plugins, APIs: what is changing?

Four protocols, four doors

Each row reads from left to right: the agent is somewhere, the protocol tells it how to talk to you, and you expose only what you choose.

MCP: plugging your product into AI assistants

What it is. Released by Anthropic in November 2024, the Model Context Protocol is an open standard that describes your tools and your data so that any assistant can use them. It is often compared to a USB-C port for AI: one connection, every assistant.

How it works. Three roles:

• The host: the AI application (Claude, ChatGPT, Cursor, VS Code…).
• The client: the connector the host creates for each server.
• The server: your program, which exposes tools (actions), resources (data) and prompts (instruction templates). It runs locally or remotely.

Where it stands.

• Adoption. In December 2025, the official blog counted more than 97 million monthly SDK downloads and 10,000 active servers. In July 2026, the main SDKs are approaching half a billion downloads a month, and the TypeScript and Python SDKs have each passed one billion cumulative downloads.
• Specification 2026-07-28. The protocol becomes stateless: each request stands on its own, with no session to maintain. A remote MCP server can now be hosted like any web API, behind an ordinary load balancer. For a small software vendor, the cost of entry drops sharply.
• Official extensions. MCP Apps (interactive interfaces displayed in the conversation), Tasks (long-running tasks) and enterprise-managed authorisation.
• Governance. On 9 December 2025, Anthropic donated MCP to the Agentic AI Foundation, a Linux Foundation fund co-founded with Block and OpenAI and backed by Google, Microsoft, AWS, Cloudflare and Bloomberg. An official server registry has existed since September 2025.

How to use it.

1. Get into the connector directories. They are the new app stores. Being there when a user asks “book me a table” or “create a visual” is worth what the top spot on Google was worth in 2010. Canva, Figma, Notion, Stripe, Zapier and Atlassian got in early.
2. Start with three high-value tools (search, create, check a status), with OAuth authentication. Write each tool's description as you would write copy: it is what the model reads to decide whether to call you.
3. Show, don't just answer. With MCP Apps, your configurator, your simulator or your quote appears in the conversation, where the decision is made.
4. Read the call logs like market research. Which tools are called, with which parameters, which ones are missing: it is a direct measure of intent. Honeycomb reports that nearly 20% of its monthly interactive queries already come from agents.
5. Automate your own growth stack. An agent connected to your MCP servers (CRM, analytics, email, Search Console) produces the weekly report, spots an anomaly and proposes an experiment. Section 15 gives a full example applied to GEO.

WebMCP: making your pages actionable by the browser's agent

What it is. A proposed web standard. The site itself declares its “tools” (search, book, request a quote), and the agent built into the browser calls them directly instead of simulating clicks and guessing what each button does. According to Chrome's documentation, this is more reliable than conventional browser control, where every step is left to the agent's interpretation.

How it works. Two APIs:

• Declarative: you add attributes to an existing HTML form. toolname and tooldescription on the form, toolparamdescription on a field to spell out what it means, toolautosubmit to allow automatic submission. Without that last one, the user clicks Submit themselves.
• Imperative: in JavaScript, you register a tool with a name, a description, an input schema and a function to run.

The tool runs in the page, in full view of the user. Your interface and your brand stay intact.

The difference from MCP. MCP runs on a server: an assistant uses you without ever visiting your site. WebMCP lives in the page: the agent has to be there already. So there is no directory of WebMCP tools, and being found remains a job for SEO and GEO (section 15).

Where it stands. Early preview in Chrome 146 in February 2026, then an origin trial from Chrome 149. For now the only implementation is Chrome's, and Google warns that the API may still change. Angular offers experimental support, and OpenAI's developer documentation includes a “Site tools (WebMCP)” page for Codex.

How to use it.

1. First annotate the forms that bring in your revenue: quote, booking, sign-up, valuation. Two attributes are enough, and it is progressive enhancement: nothing changes for humans or for other browsers.
2. Write the description like an ad. It tells the agent when to use the tool and what it will get.
3. Measure agent-driven conversions. The form's submit event carries an agentInvoked flag: you can create an “agent” segment in your analytics and compare its conversion rate with that of humans.
4. Cut the invisible failures. An agent that trips over a date picker or a five-step flow gives up without leaving a trace. A dedicated tool solves the problem.
5. Keep human confirmation for sensitive actions. No automatic submission on a payment or a signature.

Limits. It is a trial, designed for a local browser with a human present, not for headless bots. And in an iframe from another domain, tools are disabled by default unless the host site adds allow="tools": that matters for embeds (section 17).

ACP: getting your catalogue into ChatGPT

What it is. The Agentic Commerce Protocol is an open standard (Apache 2.0 licence, still in beta) maintained by OpenAI and Stripe. It was announced on 29 September 2025 together with Instant Checkout, buying inside ChatGPT, initially from US Etsy sellers.

What changed in 2026. In March 2026, OpenAI changed course. According to a spokesperson quoted by Digital Commerce 360, the priority is now product search and discovery, ACP serves as the infrastructure between users and merchants, and Instant Checkout moves into ChatGPT apps. OpenAI now describes ACP as the layer that lets ChatGPT ingest a structured catalogue, understand stock and surface the right products in the conversation. Version 2026-04-17 of the specification covers the cart, the feed, orders, authentication and integration with MCP.

So in six months, ACP went from being the engine of direct purchasing in ChatGPT to a catalogue protocol.

How it works. The merchant sends its product feed, by file or by API, along with its promotions. ChatGPT uses it to answer. Payment takes place on the merchant's site or in a ChatGPT app.

How to use it.

1. Treat the product feed like a landing page. OpenAI's guidelines cover writing product listings, modelling variants and attribution: this is merchandising, not plumbing.
2. Push your promotions through the feed, since the protocol carries them.
3. Go through your platform (Shopify, Stripe and others) rather than building it yourself.
4. If you sell a service, look at apps. The ChatGPT apps documentation describes conversion flows for booking a restaurant, requesting a quote and paying for a product. A service business therefore converts through an app, that is, an MCP server, not through ACP.

Limits. The protocol is in beta and has already changed role once.

UCP: letting agents buy from you

What it is. The Universal Commerce Protocol is an open standard launched by Google on 11 January 2026 to cover the whole buying journey, from discovery to after-sales. Its technical council brought together Google, Shopify, Etsy, Target and Wayfair; Amazon, Meta, Microsoft, Salesforce and Stripe joined on 24 April 2026, which takes it to ten companies.

How it works.

• A public profile. At /.well-known/ucp, the merchant publishes a machine-readable file that declares what it can do and which payment methods it accepts.
• Capabilities. Catalogue search and lookup, cart, checkout, identity linking (via OAuth), order tracking.
• A negotiation. The agent and the merchant keep only the capabilities they have in common.
• Several transports. REST, MCP, A2A or embedded integration: UCP builds on MCP, it does not replace it.
• The merchant remains the seller. It keeps responsibility for the sale and the customer relationship. Payment goes through “payment handlers” and, for autonomous purchases, through the mandates of the AP2 protocol.

Where it stands. March 2026: multi-item cart, real-time catalogue, identity linking. May 2026: universal cart across retailers, shoppable ads on YouTube, instalment payments in Google Pay. A specification for hotels is in draft and restaurants have been announced. Paying through UCP is still available mainly in the United States; Google has announced an international rollout.

How to use it.

1. Go through Merchant Center or your e-commerce platform, where Google is building in the activation.
2. Get prices, stock and delivery times right. The agent compares this data in real time; one wrong figure rules you out.
3. Turn on identity linking. Your members keep their loyalty benefits and you keep the customer relationship, even when the purchase happens off your site.
4. Make use of attribution. The UCP cart carries the source, the medium, the campaign and the click ID: AI surfaces can be measured like any acquisition channel.
5. Look beyond retail. Hotels and restaurants are coming, and the protocol's roles are defined by the direction of the exchange, not by the sector: it also applies to procurement between businesses.

ACP or UCP? The two coexist, and Stripe sits on both sides. My advice: do not choose, let your platform speak both.

Where to start, depending on your business

Agent Skills

Launched by Anthropic in October 2025, then published as an open standard on 18 December 2025 at agentskills.io, Agent Skills are folders containing a SKILL.md file (instructions) and resources (scripts, templates) that an agent loads on demand to carry out a specialised task. Microsoft has adopted them in VS Code and GitHub, and OpenAI uses a structurally identical architecture in ChatGPT and Codex. A directory of partner skills was launched with Atlassian, Figma, Canva, Stripe, Notion and Zapier.

According to Mahesh Murag (Anthropic), MCP provides secure connectivity to external software and data, while skills provide the procedural knowledge to use those tools effectively. For a growth hacker, a skill locks in an in-house procedure (an audit, a monthly check) so that it is carried out the same way every time.

API, connector, plugin, skill, app: who does what

Security risks

Prompt injection (an instruction hidden in a page or a document that hijacks the agent), overly broad permissions, unverified MCP servers and malicious skills are real risks: the OpenClaw case showed as much in early 2026 (section 2). The minimum rules: least privilege, human approval for any irreversible action (payment, bulk sending, deletion), servers from verified sources, and logging.

My view: MCP in 2026 is what open APIs were in 2008. The companies that expose their product to agents now will capture a disproportionate share of future usage. Those that wait will have to pay to be there. WebMCP is the cheapest bet of the four; ACP and UCP mostly concern those who sell products.

Applied to LEEZ

What it brings. Today, only people who know LEEZ look for a business there. A connector opens the listings to anyone who hands their search to an assistant, and to every fiduciary (a Swiss accounting and trust firm) that wants to check, from its own tool, whether a client could find a buyer. The more a listing is seen by serious buyers, the more likely the transfer becomes.

In practice.

• First, WebMCP on the valuation and alert forms, at low cost.
• Next, an MCP server with three tools: search for a listing, start a valuation, create an alert.
• ACP and UCP remain beside the point: you do not buy an SME the way you fill a cart.

Opening a door to agents is pointless if none of them comes knocking. That leaves the most hotly contested question of 2026: how does an AI choose who it cites?

↑ Contents

Section 15

SEO and GEO in 2026

Where the data in this section comes from

This section draws on two posts by Nicholas Dulait, co-founder of ChatSEO, published on X in 2026 (links at the end of the guide). They report on the panel run by Jacky, of the French link-building platform Links Garden: the same million prompts sent to ChatGPT every month for five months, which makes 5 million answers from which every brand and every source is extracted. This data was discussed on the MVP podcast with Paul Vengeons (co-founder of ChatSEO).

Three caveats are in order:

• This is French data, on ChatGPT alone; collection on Gemini has only just started, and Claude or Gemini may weight things differently.
• The panel is private and unpublished: I cannot verify it independently.
• It is reported by people who sell SEO tools and links, and who have a commercial interest in these conclusions.

Nicholas Dulait himself acknowledges that the patterns shift and that part of his analysis will be out of date within months. So I treat these figures as the best measurement available, not as established truth.

The state of SEO in 2026

Zero-click is becoming the norm. The Pew Research Center study (July 2025, 900 US adults, 68,879 real searches) shows that users click on a traditional result in 8% of visits when an AI summary is displayed, against 15% without one; they click on a link cited in the AI summary in only 1% of cases, and end their session more often (26% against 16%).

The decline is accelerating. In April 2025, Ahrefs measured a CTR for position 1 that was 34.5% lower when an AI Overview was present; its February 2026 update, covering 300,000 keywords, measures a 58% drop. An important nuance, underlined in the same study: the CTR of informational queries without an AI Overview has also fallen sharply (from 7.6% to 3.9%), a sign of a wider change in behaviour.

France's turn has come. The organisers of the SEO & GEO Summit (Disneyland Paris, 15-16 October 2026) point to the rollout of AI Overviews in France in summer 2026. For a French-language site, the impact on traffic is therefore under way, not behind us.

GEO: being cited rather than ranked

Generative engine optimisation (GEO) aims to get you cited as a source in the answers of ChatGPT, Claude, Perplexity, Gemini, Copilot and AI Overviews. Answer engine optimisation (AEO) is a variant focused on direct answers. The metric changes: you move from click-through rate to reference rate, the share of generative answers, across a panel of queries, that mention or cite your brand.

How ChatGPT chooses its sources

Fan-out: the real keyword research of GEO

When ChatGPT decides to search, it does not search for the prompt: it rewrites it as two to five short Google queries. The example Jacky gives: someone explains that they hurt their back at judo, saw their doctor and have slept badly ever since; the model searches for neither judo nor the doctor, it searches for the best mattress for back pain, with the year.

People often object that every prompt is unique. That is true, but the queries the model derives from it, the fan-outs, converge. Thirty people write thirty different prompts, and the model boils them down to the same three searches. The GEO keyword list is the list of these short queries, and it looks a lot like classic SEO.

The word “best” is there because the user wants a recommendation; the year, because the model wants fresh sources.

Reddit: a lever that is fading

Betting on Reddit was the GEO reflex of 2025; the panel data contradicts it. At the end of 2025, Reddit appeared in about 80% of ChatGPT's answers. In the French data, its share of citations went from about 30% in spring 2026 to 9% in July, then to around 5% in August, and it is still falling. In some sectors it never worked: in car insurance, Reddit has been stuck at 3% for five months.

The rule behind these figures: any source that becomes too effective gets spammed by SEOs, then the model demotes it. A GEO strategy that comes down to posting on Reddit is optimising for December 2025.

Better to pick three or four networks, the ones your buyers use:

• Cosmetics, food, lifestyle: Instagram and TikTok.
• Services and B2B: LinkedIn, X, Reddit.
• Software: X, YouTube, GitHub, Reddit.

Run them like a real account: most posts serve the audience, and one in three to one in five targets a query the model searches for.

Stop ranking yourself first

Putting yourself at number one in your own “best of” ranking was the cheapest GEO trick; it has become the surest way to be ignored. Paul Vengeons did it for years on his consultancy site, and ChatGPT cited him with his own page as the source. He is still cited, but the source is now the ranking published by another company, Convertix. An infographic from the agency Eskimoz sums up the mechanism: when a site ranks itself first, the model tends to cite numbers two and three and to skip number one.

It is a trade-off, not a rule:

• On classic Google, a self-ranked page can still win.
• For GEO, it no longer gets you named.
• With the reader, it costs trust, and in some professions it is ruled out (a pharmacy cannot proclaim itself the best).

The method ChatSEO has settled on: leave alone the pages where you are first and that get clicks, and rewrite on merit the ones that get none, one page at a time. A rewrite by an agent is expensive in usage and can delete tables without any visible error: compare the old and new versions before publishing.

Being ranked by third parties: the swap and the ladder

One ranking on someone else's site is worth more than ten on your own, for the model and for the reader alike. A restaurant that calls itself the best convinces no one; a local guide that names it first does.

The ranking swap. Paul Vengeons lists agencies in his rankings and asks in return for a ranking on their site that cites him. Two rules make the effect last: freshness (redo the swap every year, with the new year in the title) and volume (most companies reportedly have at least 20 GEO queries that deserve a swap).

The placement ladder, from the cheapest to the most demanding:

The expensive citation is no better. According to Jacky, a €1,000 or €2,000 mention in a major outlet is cited for one to three months and then drops out of the rotation; a €10 or €20 placement on a small, relevant site does the same. His network sites appear in about 30% of answers, and so do the sites of “real” publishers. The premium for a major outlet mostly pays for a logo.

Reviews, Wikipedia and meeting people

• Customer reviews: one platform, not three. Local service: a Google Business Profile listing. SaaS and e-commerce: Trustpilot. Scattered reviews make three weak profiles instead of one strong one.
• Wikipedia: do not ask. The moderators know why marketers turn up; you get in when someone else decides you belong there.
• Meeting people: a message to someone you met at an event is worth more than any cold email. Go to your niche's events with the list of the five sites where you want to appear (see section 22).

My reservations about these tactics

I am adopting the measurement, not all the methods. Four limits:

1. A swapped or paid ranking is still commercial communication. Presented as an independent opinion, it can amount to misleading information under Art. 3 of the Swiss Unfair Competition Act (UCA), and European law on unfair commercial practices targets advertising disguised as editorial content. Disclose the partnership, and have a lawyer check it.
2. Bought links and large-scale link exchanges breach Google's spam policies. And the model still opens Google: a Google penalty costs you the citation as well.
3. The outreach Nicholas Dulait describes relies on mailboxes created on a domain that looks like yours, then “warmed up”. It is a set-up designed to protect the reputation of the main domain, and in Switzerland it runs up against Art. 3 para. 1 let. o UCA (see section 21). In French-speaking Switzerland, five personal emails a day from your real address, or meeting in person, work better.
4. The model has learned to ignore self-ranking; it will learn to ignore rankings handed out as favours. What does hold more generally, and Nicholas Dulait says so himself, is that a third party speaking well of you is worth more than you doing it yourself. You might as well earn it: original data, real customers, real reviews.

Measuring: staying in the rotation

GEO is not a rank you hold, it is a rotation you stay in. The minimum monthly routine:

1. For each fan-out query, record the Google top 10 and the rankings or comparisons that appear in it.
2. Check that you are still listed in the third-party rankings, and in what position.
3. Spot the titles that still carry last year's date and suggest an update.
4. Test ten queries by hand in ChatGPT, logged out, and note: cited or not, through which source.

The fourth point stays manual on purpose: the ChatGPT API does not return the same sources as the web app, so a dashboard built on the API does not measure the same thing. On the analytics side, also segment the referral traffic coming from chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com.

Automating GEO with agents

Nicholas Dulait describes a set-up in which Claude Code plays the role of “Chief of GEO”: the agent plans, writes, finds prospects and prepares follow-ups; ChatSEO, connected through MCP, acts as the analyst (Search Console, volumes, live Google results); a sending tool dispatches the emails; the human approves anything that touches a live page or goes out under their name. The procedures are Agent Skills (section 14), four of which are detailed in his posts: building the query list, auditing self-rankings, preparing the outreach, running the monthly check.

Three safeguards from these skills are worth copying as they stand into any growth automation:

• Never invent a volume or a ranking: an empty cell is better than a plausible number.
• List everything that disappears between the old and the new version of a page, and publish nothing without approval.
• Never guess an email address, and have every reply approved before it is sent.

Mistakes to avoid

• Tracking prompts rather than fan-outs.
• Building your strategy on Reddit, or on any single source that SEOs are busy saturating.
• Ranking yourself first in your own rankings.
• Paying the premium for a major outlet in the hope of a lasting citation.
• Buying a tool for its sentiment score.
• Mass-producing generic AI content: it will be neither ranked nor cited.
• Blocking AI search crawlers by mistake while trying to block training crawlers.
• Believing that an llms.txt file is enough.

Applied to LEEZ

What it brings. A 62-year-old owner who asks an AI how to sell their business should get a reliable answer, not silence. Every time LEEZ is cited, one more entrepreneur finds out that an organised exit exists, early enough to prepare for it.

In practice.

1. Target the short queries, in all four languages: “meilleure plateforme vente entreprise Suisse”, “Firma verkaufen Plattform”, “LEEZ avis”.
2. Publish no comparison in which LEEZ ranks itself first. Earn its place in the lists kept by third parties.
3. Publish data that no one else has: the business transfer barometer. Every quarter, it would give the active listings by canton and by sector, the asking prices and the ratio of buyers to listings. LEEZ's 2026 guide to acquisition financing, with its ranking of the 26 cantons, is a first asset of this kind.
4. Look after the profiles and the reviews on the platforms that AI assistants read.
5. Shoot one video per seller question, and publish the transcript on the site.
6. Date every page, as the listing counters already do.
7. Run a monthly check on twenty queries per language, noting who is cited.

The method is set out. What do Google and AI assistants answer, today, when you look for a business transfer platform in Switzerland? I ran the test, query by query.

↑ Contents

Section 16

Worked example: SEO/GEO

What LEEZ shows publicly

I reread leez.ch the way a crawler would, adding nothing to what is visible.

LEEZ already has what AI assistants like to cite: hard figures that are fresh and broken down by canton and by sector. What it lacks is third parties to say so in its place.

The queries to target

Who comes up today

These readings show a trend. The exact positions, by language and by canton, are in Search Console.

The lesson: in French, there is no Swiss comparison of business transfer platforms. Someone will fill that gap. Better that it be a credible third party, fed with LEEZ's data, than a competitor.

If no third party cites LEEZ yet, what remains is to go where the trusted third parties already are. An old hack makes that possible: installing a piece of your product on their sites.

↑ Contents

Section 17

The embed hack

The mechanics: the product lives on other people's sites

Each installation is at once a distribution channel, social proof and a brand mention.

The technical formats

What the embed really does for SEO

What works. Content served as HTML on the partner's site: Google's documentation points out that server-side rendering is still preferable, because not all crawlers run JavaScript. On top of that come the brand mention and the referral traffic, even with a nofollow link.

What works poorly. The iframe as a vehicle for content you want ranked. Google publishes no general rule, but its consistent advice is to include directly in the page whatever you want indexed. My reading: an iframe passes no useful authority to the embed provider and does not reliably enrich the partner's page.

What is risky.

• Optimised widget links. Google's spam policies explicitly target keyword-rich or hidden links embedded in widgets distributed across many sites.
• The link imposed by contract. Google also counts as spam the act of requiring a link without leaving the site owner the choice of qualifying it.
• The link in exchange for a commission. A link obtained in exchange for a benefit must carry the nofollow or sponsored attribute.
• Duplicate content. If fifty partners republish the same listings in full, you create fifty competing copies of your own pages.

My attribution link. A visible credit whose anchor text is the brand and never a keyword. The partner can edit it or remove it. It is marked sponsored as soon as a referral commission exists. I knowingly give up the authority of the widget link: the risk of a penalty outweighs the gain, and the GEO effect does not depend on that attribute.

What AI assistants see

The Vercel and MERJ study, published in December 2024, remains the broadest public measurement: it observed no JavaScript execution by the crawlers of ChatGPT, Claude or Perplexity. Gemini relies on Googlebot's infrastructure, which does run scripts. The measurement is getting old, but as of mid-2026 no AI vendor has documented a change. So I work on the assumption that a JavaScript or iframe widget is invisible to these three assistants.

Three consequences:

1. The block must exist as static HTML on the partner's site, readable without a script.
2. Repeated mentions do the work. Fifty partner pages saying, each in its own words, that they work with you: that is exactly what section 15 describes as decisive, being recommended by a third party.
3. The partner's page can itself rank for a local query. That is more than a link: one more third-party page among the sources the model rotates.

The legal framework

Distributing through other people's sites does not excuse you from keeping your own house in order. You still need to know what is wrong with it.

↑ Contents

Section 18

Auditing your site, and cleaning up your backlinks

What you measure

Three ways to do it with Claude

1. In the conversation, with your exports. You attach the Search Console export, the PageSpeed report and the list of links, and you ask for a ranking. No access to the site is needed. The limit: Claude only analyses what you give it.
2. With Claude Code and a browser it controls. The Chrome DevTools MCP server, maintained by the Chrome team, gives the agent a real browser: Lighthouse audit, performance trace, Core Web Vitals, console, network. It works with Claude Code and with other agents alike. You give it ten landing pages, it measures them and suggests the fixes.
3. As a routine, with a skill. You write the procedure once in an Agent Skill (section 14), modelled on the monthly check in section 15: same pages, same measurements, comparison with the previous month, five ranked actions.

The instructions that make a good audit

• Give the context: the type of site, the pages that bring in revenue, the goal for the quarter.
• Demand evidence: every finding with its URL and its measurement, and an empty cell rather than a plausible number.
• Ask for a ranking by expected gain and by effort, not a list of eighty items.
• Separate the finding from the fix: the audit changes nothing. Every change goes through approval, page by page, with the list of what disappears.
• Have it measure again after the fix.

Three requests that work:

• “Here is the Search Console export for the last sixteen months. Rank the twenty pages that lost the most clicks, and say for each one whether the likely cause is position, click-through rate or demand.”
• “Audit these ten pages on mobile. For each one, give the measured LCP, INP and CLS, the main cause and the fix with the best gain-to-effort ratio.”
• “Open these five pages without JavaScript and tell me what an AI crawler can still read on them.”

The limits

• Lab versus field. Lighthouse measures in the lab, whereas Google judges on real visits. The two numbers differ, and the second is the one that counts.
• A suggested cause remains a hypothesis until the follow-up measurement has confirmed it.
• Access. Read-only on Search Console, no password in the conversation. The Chrome DevTools server exposes to the agent everything the browser displays: I give it a separate browser profile of its own.

Cleaning up bad backlinks: almost never

Google ignores most spam links on its own, and its spokespeople keep saying so: the vast majority of sites never need the disavow tool, which is still available in 2026. “Toxic link” is not a Google term, and the toxicity scores in third-party tools do not reflect its algorithm. An agency that sells link clean-up as a maintenance service is mostly selling fear.

A disavow is justified in two cases:

• A manual action for unnatural links appears in Search Console.
• You know that links were bought or manufactured for the site, by you or by a former provider, and you want to get ahead of the problem.

In both cases, here is the procedure:

1. Check the manual actions section in Search Console. If there is nothing there and no history, stop here.
2. Export the links from Search Console and a third-party tool.
3. Have Claude sort them, domain by domain: known bought or exchanged link, site network, directory, automated spam, legitimate link. A human reviews every domain before any decision.
4. First ask for the removal of the links you created yourself.
5. Submit the disavow file, one line per domain, in Google's tool, and keep a dated copy: each new file replaces the previous one.
6. Request a reconsideration if there was a manual action.

The risk is real: disavowing a good link means asking Google to ignore it. The tool does more harm than good when you use it as a precaution.

The clean-up that pays off lies elsewhere: redirecting deleted pages that still receive links, repairing broken internal links, removing redirect chains, correcting outdated directory entries. It is also the simplest argument against the bought links in section 15: what you do not buy, you do not have to clean up.

Applied to LEEZ

What it brings. A listing that Google does not index, or a page that takes six seconds to load on a phone, is a business its buyer will never see. The audit does not serve the site: it serves every business published on it.

In practice. On a listings platform, four points deserve a regular check: listings that a business broker also publishes on other platforms and which page Google should keep, what happens to expired listings, consistency across the four languages, and the speed of the list pages, the most viewed on mobile.

The agent measures, sorts and suggests. But to direct it and review its work, do you still need to know how to code?

↑ Contents

Section 19

Programming languages: a comparative study in the age of vibe coding

The context in 2025–2026

According to GitHub's Octoverse 2025 report, TypeScript became the most used language on GitHub for the first time in August 2025, with 2,636,006 monthly contributors (+66% year on year), about 42,000 more than Python. GitHub attributes this growth to AI tools, which produce more reliable code with strict types, and notes that Python remains dominant in AI and machine learning. The same report says that 80% of new developers use Copilot in their first week.

Comparison table for a growth hacker

Final recommendation

1. SQL first: it is the language of truth. Without SQL, you depend on someone else to know whether your test worked.
2. TypeScript next: the best choice for building websites, MCP servers, apps inside assistants and deployed automations. Typing makes vibe coding safer.
3. Python as a complement: for data analysis, scraping and agent prototypes.
4. n8n or Make to orchestrate, without coding, whatever needs orchestrating.
5. Go and Rust: only if you are building infrastructure. Bash: enough to read and run scripts.

Applied to LEEZ

What it brings. To help an owner hand over a business, you need to know where the market gets stuck: in which canton and in which trade buyers are in short supply. Those answers are in the data, provided you know how to query it.

In practice.

• SQL first: the number of buyers per listing in each canton and each sector, the time to first contact and the number of introductions per week can all be calculated in a few queries.
• TypeScript next: it is the language of an MCP server.
• Python for the barometer: cleaning the listings, deduplicating those a business broker has published several times, producing the quarterly series.

Knowing how to query data assumes you have some. The temptation then is to go and take it from others, and that is where the law comes in.

↑ Contents

Section 20

Scraping: good practice in the age of AI

The legal framework

Switzerland. The new Federal Act on Data Protection (FADP), in force since 1 September 2023, applies as soon as personal data is collected, even if it is public; the principles of transparency, proportionality and purpose limitation apply, and the criminal fines (up to 250,000 Swiss francs) target the individuals responsible. The Unfair Competition Act (UCA) can penalise taking over someone else's work product by technical means without an appropriate effort of one's own (Art. 5 let. c UCA). Copyright protects original content.

European Union. The GDPR applies to personal data collected by scraping. The sui generis database right protects substantial investments. The text and data mining (TDM) exception in Directive 2019/790 allows extraction, unless the rightholder has opted out in a machine-readable way (for example via robots.txt or metadata).

Notable recent case law.

• GEMA v OpenAI (Munich Regional Court I, 11 November 2025, 42 O 14139/24): the court held that the memorisation of song lyrics in the models and their reproduction in the answers constitute copyright infringements, and that the TDM exception does not cover this situation. The judgment is not final.
• Getty Images v Stability AI (English High Court, 4 November 2025): Stability AI largely won on copyright, the court finding that the model neither stores nor reproduces Getty's works, with a limited trade mark infringement linked to the watermarks; Getty had dropped its main claim on training.

How to read this: continental Europe is proving stricter than the United Kingdom. For a growth hacker, the question is not just “do I have the right to collect?” but “what am I going to do with the data, and am I reproducing protected content?”

The web is closing to AI crawlers

On 1 July 2025, Cloudflare began blocking AI crawlers by default on new domains and launched “Pay per Crawl”, which lets sites charge for access via the HTTP 402 status code. Its June 2025 data showed crawl-to-referral ratios of about 14 to 1 for Google, 1,700 to 1 for OpenAI and 73,000 to 1 for Anthropic. On 1 July 2026, Cloudflare announced the next step, applied since 15 September: blocking AI crawlers by default on advertising-funded pages. According to several secondary sources, Pay per Crawl is also due to give way to a “Pay per Use” model, which pays publishers when their content is used in an answer. The settings that follow from this are described in section 28.

Good practice

1. Prefer the official API or a licensed data feed whenever one exists.
2. Respect robots.txt and the terms of use; identify your bot with an honest user agent and contact details.
3. Limit your request rate so that you do not degrade the service you are targeting.
4. Minimise personal data: collect only what is necessary, document the legal basis, and inform people if you contact them afterwards.
5. Never bypass authentication, a paywall or a technical block.
6. Do not reproduce protected content; extract facts, not works.
7. With LLMs: AI-assisted extraction (tools that turn a page into structured data) is powerful, but the same rules apply; agentic browsers that act “like a human” do not exempt you from the terms of use.

Alternatives

Open data (opendata.swiss, commercial registers via Zefix), data-sharing partnerships, proprietary surveys, buying licensed datasets, and zero-party data declared by your users.

Applied to LEEZ

What it brings. The more widely a listing circulates, the better the chances that the business finds a buyer. But the owner's identity must never circulate with it. The whole challenge is to let the one through without letting the other leak.

In practice. For a marketplace, the question arises mostly in one direction: the one where others come to take its listings.

• When it is scraped, its listings feed aggregators and AI answers that send no one back. The defence: a public excerpt, the detail behind the non-disclosure agreement (NDA), and clear terms of use.
• Better to offer a feed than to be plundered: an official, tracked feed (section 35).
• When it collects data in turn, the rule is the one that runs through this whole section: open sources or partners who have given their consent, the origin stated, and never a competitor's work taken as it stands (Art. 5 let. c UCA).

Other people's data can only be taken under conditions. Yet there is one channel whose data belongs entirely to you, and which is pronounced dead every year: email.

↑ Contents

Section 21

Email marketing: automation, newsletter, deliverability

The state of play in 2026

Two forces have transformed email: the authentication requirements of the major providers and AI summaries in inboxes.

Deliverability: the rules of the game

• Google and Yahoo have required the following since February 2024 from anyone sending more than 5,000 messages a day: SPF, DKIM, DMARC, one-click unsubscribe and a low complaint rate.
• Microsoft has applied similar requirements since 5 May 2025 for Outlook.com, Hotmail.com and Live.com: SPF and DKIM must pass, and a DMARC record of at least p=none, aligned with SPF or DKIM, is required for anyone sending more than 5,000 emails a day; non-compliant messages are rejected.

My advice: set up SPF, DKIM and DMARC even if you send 50 emails a week. It costs nothing and has become a condition for existing at all.

AI summaries in inboxes

Gmail, Apple Mail and Outlook now offer AI-generated summaries. The consequences: your subject line and your first sentence must carry the essential message; a clear, structured email will be summarised better; open rates become even less reliable as a metric, so favour clicks, replies and conversions.

The newsletter as owned media

A newsletter is an asset that depends on neither an algorithm nor a search engine. In 2026, it is also a source of original content that you can republish on your site to feed your GEO.

Automation sequences that work

1. Welcome (3 to 5 emails): promise, best content, first action.
2. Activation: triggered by behaviour (no usage after 3 days).
3. Re-engagement: before you delete inactive subscribers (which protects your deliverability).
4. Enhanced transactional: confirmations and summaries that add value.

Cold emailing: the legal framework

• Switzerland: Art. 3 para. 1 let. o of the Unfair Competition Act (UCA) prohibits sending mass advertising by electronic means without prior consent, without correctly identifying the sender or without a simple, free way to opt out; an exception exists for existing customers in respect of similar products. The Federal Act on Data Protection (FADP) applies to the processing of addresses.
• European Union: the ePrivacy Directive requires prior consent for natural persons; B2B rules vary from country to country (France, for example, tolerates B2B prospecting to a professional address if the message relates to the recipient's role and a right to object is offered). The GDPR applies to all personal data.

My view: in 2026, mass cold emailing is a losing bet (legally risky and technically filtered). Highly targeted, personalised, low-volume cold emailing remains viable in B2B.

Tools

Brevo (French in origin, suited to the GDPR context), Mailchimp, Klaviyo (e-commerce), Customer.io (product), Beehiiv and Substack (media newsletters), as well as solutions hosted in Switzerland or the EU for sensitive data.

Applied to LEEZ

What it brings. A business transfer takes years to prepare and is settled in a few weeks. Email is the only channel that lasts that long: it stays with the owner for as long as they hesitate, and it alerts the buyer the day the business they were waiting for appears.

In practice. Three audiences, so three lists and three tones.

• For buyers, an alert by criteria (canton, sector, budget) and a weekly digest.
• For sellers, a slow sequence after the valuation. The decision to sell matures over months: pointers to help them prepare, never a pushy follow-up.
• For partners, a monthly letter on the new deals in each canton.
• What LEEZ does not do: write en masse to company directors whose address comes from a register. The UCA regulates it, and the platform's reputation would not recover.

Email maintains a relationship; it does not create one. To create one, nothing has replaced what machines cannot do: bring people together in the same room.

↑ Contents

Section 22

Building a community: the need for humans to meet

The data on loneliness

The report of the WHO Commission on Social Connection (30 June 2025) estimates that one person in six worldwide suffers from loneliness, and that loneliness is associated with about 871,000 deaths a year, or about 100 deaths an hour. Adolescents aged 13 to 17 are the most affected (20.9%). In May 2025, the World Health Assembly adopted its first resolution on social connection.

What this means for a growth hacker: meeting in person is becoming rare, and therefore valuable. It is a competitive advantage that AI cannot copy.

The return of in-person events

Run clubs, dinners with strangers, themed meetups, founder communities: the format matters less than three constants, which are a regular ritual (same day, same place), a light but real barrier to entry (registration, a small contribution) and a role for each member.

The mechanics of community-led growth

1. Core: 10 to 20 committed people who keep coming back.
2. Ritual: a recurring meeting and a recognisable format.
3. Identity: a name, a symbol, a shared language.
4. Contribution: members create content, organise, invite.
5. Amplification: each event produces content (photos, stories, testimonials) that recruits the next attendees.

The Burgerpass (section 4) is a local example: a community of “burgerlovers” that brings consumers and restaurant owners closer together.

My advice

Do not launch a “community” on Discord and hope it will live on its own. Start with a monthly in-person event for 20 people; the online community will follow.

Applied to LEEZ

What it brings. Nobody hands their business over to a stranger. The decision is made between two people who have met and who trust each other. Creating these meetings shortens the path between an owner who wants to leave and someone who wants to carry on.

In practice. The two audiences do not relate to visibility in the same way: the buyer wants to be seen, the seller wants anything but.

• A buyers' circle in each canton, with open meetings.
• Closed meetings for sellers, by invitation, organised with their fiduciary (a Swiss accounting and trust firm).
• Evenings with the fiduciary associations of each canton, where LEEZ presents its figures on business transfers.

It is also the best answer to disintermediation: people come back to the platform for what it brings together, not just for a listing.

Twenty people in a room is a start. How many does it take to keep a business alive? Fewer than you might think, and a 2008 essay gave the figure.

↑ Contents

Section 23

The 1,000 true fans

Kevin Kelly's idea

In 2008, Kevin Kelly, co-founder of Wired magazine and its first executive editor, published an essay that became a classic: “1,000 True Fans”. His thesis: to make a living from their work, a creator does not need a mass hit. A thousand true fans are enough, meaning people who buy everything the creator produces. At $100 of margin per fan per year, that makes $100,000 of income.

Kelly sets two conditions:

• Produce enough each year for every fan to have something to spend that sum on. It is easier to give more to your current customers than to find new ones.
• Keep a direct relationship, with no intermediary taking most of the money or holding the contact in your place.

The figure is an order of magnitude, not a law. If there are two of you, you need two thousand fans. At $50 a year, you also need two thousand. Around this core orbit circles of occasional fans: you work only for the centre, and the rest follows.

What has changed since

In 2020, Li Jin, then at Andreessen Horowitz, proposed dividing the number by ten: a hundred people willing to pay $1,000 a year. Kelly welcomed this extension. The difference is not just arithmetic: at that price, you are no longer selling patronage, you are selling a result.

Three things have shifted since 2008:

• The tools for a direct relationship exist: paid newsletter, subscription, private community.
• Platforms have stepped in between. An algorithm decides who sees what, and an AI sometimes answers in your place. Kelly's second condition has become the hardest to meet, hence the importance of email (section 21) and of meeting in person (section 22).
• Content has become abundant. When everyone can produce, people grow attached to people, not to texts.

The fairest criticism: a thousand true fans is a lot. Few people buy everything someone produces, and it takes years to gather them. Nor does the theory hold for a one-off purchase: you do not become a fan of a product you buy only once.

How I use it

1. Define the true fan by a measurable behaviour, not by a feeling.
2. Count them honestly. The number is always smaller than you think.
3. Know the first hundred by name and talk to them directly.
4. Keep a channel that no platform can cut off.
5. Give them a reason every year to buy, to come back or to recommend.
6. Track the share of the business that comes from them.

Applied to LEEZ

What it brings. The fiduciary (a Swiss accounting and trust firm) sees an owner's retirement coming years before anyone else: it shows in the accounts. If the fiduciary thinks of LEEZ at that moment, the transfer is prepared in time instead of being improvised.

In practice. A marketplace upsets the theory. The seller sells their business only once: they will never be a fan in Kelly's sense. The true fans are elsewhere.

• The fiduciaries and business brokers who bring in every deal. This is the hundred-true-fans model: a hundred active partners are worth more than ten thousand visitors.
• Buyers on active watch, who open every alert and sign several non-disclosure agreements (NDAs). The site shows just over 300 registered buyers: I would aim for a thousand verified and genuinely active buyers.
• Former sellers and buyers who agree to give a testimonial.

Two measurable definitions: a partner who has brought in at least three deals in twelve months, and a buyer who gets at least one qualified introduction per quarter. The true fan is also the best answer to disintermediation: they come back to the platform even though they could do without it.

Fans like these deserve better than a campaign. There is one kind of message that almost no one sends any more, and that no one throws away unread.

↑ Contents

Section 24

The letterbox: writing by hand, a forgotten practice

Why it still works

A physical letterbox is less crowded than an email inbox; a handwritten letter is perceived as an effort, and therefore as a mark of respect. It is the principle of reciprocity in its purest form.

Response data

Editions of the ANA/DMA Response Rate Report in the United States, quoted by secondary sources, give response rates for physical mail of 5 to 9% on house lists and 3 to 5% on prospect lists (2018 edition), and an average rate of 4.4% for mail against 0.12% for email. These figures are old and American: the best approach is still to measure your own rates with a dedicated code or address.

The formats

• Genuine handwritten letter: for your 50 most important prospects or customers.
• Robot-written “handwritten” letter: some services use robots holding real pens to produce letters that look handwritten, at scale. Effective if the text is truly personalised; counterproductive if the recipient discovers that the intimacy is fake.
• Printed direct mail: cards, catalogues, samples.

The Swiss context

Swiss Post offers addressed and unaddressed advertising mail services (PromoPost). Be careful: many households display the “No advertising” sticker, which must be respected for unaddressed mailings. For addressed mailings, the Federal Act on Data Protection (FADP) applies to the use of addresses: they must come from a lawful source and the recipient must be able to object.

My advice

Combine them: an announcement email, a handwritten letter for key accounts, a follow-up phone call. For a B2B SME in French-speaking Switzerland, 20 handwritten letters a month to targeted decision-makers are often worth more than 2,000 cold emails.

Applied to LEEZ

What it brings. Many owners close to retirement are not searching for anything online. They have not yet decided to sell, and no online advertising will reach them. A letter finds them where they are, while there is still time to prepare what comes next.

In practice. Twenty to fifty letters a month, on a narrow segment, for example the restaurant owners of one canton: a handwritten card, a sentence about discretion, a QR code to the free valuation. The addresses come from lawful sources, each letter offers a simple way to stop receiving them, and a QR code unique to each mailing makes it possible to measure the results.

The letter reaches fifty people. To reach fifty thousand in a region, a channel once written off is making a comeback: regional radio.

↑ Contents

Section 25

Radio advertising and recycling it as a podcast

Audiences in Switzerland

According to Mediapulse, in the second half of 2025, people aged 15 and over listened to an average of 76 minutes of live radio a day, an audience that held firm despite the partial FM switch-off. Daily reach stood at 76% in Italian-speaking Switzerland, 70% in German-speaking Switzerland and 65% in French-speaking Switzerland. In the first half of 2026, radio had more than 5 million daily listeners and nearly 7 million weekly listeners; in French-speaking Switzerland, 64% of residents listen to it every day, for an average of 88 minutes per listener.

The effect of the FM switch-off. SRG SSR gave up FM at the end of 2024. According to figures quoted by private stations, RTS lost nearly a quarter of its listeners in the first half of 2025 (-19% for La Première, -46% for Couleur 3, -49% for Espace 2), to the benefit of foreign and private stations in particular. One FM claimed 341,000 weekly listeners in the first half of 2025, its best audience in four years, and Radio Lac 180,000. According to SRG SSR, more than 80% of radio minutes are now listened to digitally (internet or DAB+).

The growth angle. Regional private stations in French-speaking Switzerland are gaining listeners and remain affordable: an underrated channel for a local brand. For rates, ask the advertising sales houses directly.

Effectiveness

Radio is a medium of repetition and context (car, morning, work). It excels at local awareness, time-limited offers and simple messages. It is hard to measure through direct attribution: use a dedicated short URL, a promo code or a specific phone number, and watch brand searches during the campaign.

Podcasts

Two formats dominate: ads read by the host (host-read), perceived as a recommendation, and dynamic ad insertion (DAI), which lets you target spots and replace them across the back catalogue of episodes.

Recycling radio into podcasts and vice versa

1. Radio spot → podcast pre-roll: same script, adapted to be read by the host.
2. Radio interview → podcast episode: secure the rights, publish the long version, cut it into video clips for social media.
3. Podcast → radio segment: offer a local station a 2-minute segment drawn from your episodes.
4. Everything → text: transcribe, structure and publish on your site for SEO and GEO.

Applied to LEEZ

What it brings. The tradesperson or restaurant owner approaching retirement listens to regional radio in the car or in the workshop. Telling them there, in plain words, that selling is possible and discreet shows them a door they did not know to look for.

In practice. The message fits in one sentence: “Thinking of handing over your business? Value it in five minutes, in complete confidence.” A short URL per station makes it possible to measure the results. A monthly segment, “a business to take over”, then becomes a podcast episode, then a page on the site.

A spot, a podcast, a page: keep producing content and one question eventually comes up. Should you go all the way and become a media outlet yourself?

↑ Contents

Section 26

Why create a company media outlet

The question

Should you hire a writer, fit out a media room, produce your own videos and put your name to editorials? In other words, is a company better off behaving like a media outlet than leaving its content to its marketing department?

I am not talking about launching an independent media outlet that would live off its subscribers or its advertising. I am talking about a media outlet funded by a company, to serve its business.

What separates a media outlet from a marketing department

The test comes down to one question: if the company stopped selling tomorrow, would the audience keep reading? If so, it is a media outlet.

The idea is more than a century old

• John Deere has published The Furrow, a magazine for farmers, since 1895.
• Michelin has published, since 1900, a guide that gives motorists reasons to drive, and therefore to wear out tyres.
• Coop and Migros publish the weeklies that still top the list of Switzerland's most-read print titles.
• Red Bull founded Red Bull Media House in 2007, a production company in its own right.
• HubSpot bought the newsletter The Hustle in 2021, rather than go on renting other people's audiences.

On the scale of an SME, that is what we did with Tayo Tayo (section 6): a media outlet for the property world, filmed right inside our clients' offices.

Why the question comes up again in 2026

• Generic content is no longer worth anything. An AI writes a decent article on any subject in a few seconds. What keeps its value is what an AI cannot manufacture: your own data, access to the field, faces, an opinion you stand behind.
• Machines cite sources. A media outlet that produces original figures and analysis becomes a source for AI assistants (section 15).
• An owned audience protects you from algorithms. A subscriber list depends on neither a search engine nor a social network.
• Production costs less. A quiet room, two cameras and good microphones are enough to keep a regular format going.

What you need to put in place

When it is better not to

• Without material of your own. Without data, access to the field or rare expertise, the outlet will produce nothing but generic content.
• Without patience. Results take 12 to 24 months.
• With a management team that wants to review everything. If every article has to sell, the team goes back to being a marketing department and the audience leaves.
• If you are counting on traffic. With zero-click answers, you have to aim for awareness and citations, not visits.
• If you are counting on budget. The American fund Andreessen Horowitz launched its media outlet, Future, in 2021 and shut it down the following year.

My advice

Run a 90-day test before hiring anyone: one narrow topic, one format, one regular slot, one byline, and outsourced production. If subscribers stay and reply, hire the editor-in-chief. The media room comes last.

Stay narrow and deep: “growth hacking in French-speaking Switzerland” rather than “digital marketing”. Three formats are enough: a flagship format (the annual report), a recurring format (the monthly newsletter), a human format (meetups or a podcast).

Applied to LEEZ

What it brings. Business transfers are seldom talked about. Owners do not dare, and many young people do not know that taking over a business is another way of being an entrepreneur. A media outlet makes the subject visible, with figures that show where SMEs are at risk of disappearing.

In practice. The quarterly barometer is the flagship format, the monthly newsletter the recurring format, the meetups the human format. The reference guide to acquisition financing, in its 2026 edition, is another flagship format. Other players already publish a newsletter on business transfers in French-speaking Switzerland: LEEZ stands out through what it alone has, the data from its listings.

With what resources. No media room to start with. One person responsible for the editorial line, outsourced video production, and a regular slot kept up for a year before investing in a studio.

A media outlet is built over two years. At the other end of the spectrum is something that plays out in forty-eight hours: the meme.

↑ Contents

Section 27

Memes and viral videos

What recent cases show

The figures that circulate on the effectiveness of memes (60% engagement, 20% click-through rate) come from marketing blogs with no verifiable method. I disregard them.

The rules

1. Reuse a format that already exists. A meme is recognised in one second. You do not invent a format, you slip your subject into it.
2. Move fast. The window lasts one to two days. You need an approval process decided in advance.
3. Put the product in the joke, not next to it.
4. Bet on a series rather than a one-off: same format, same day, same voice.
5. Polish the first three seconds of a video, with subtitles and a face.
6. Measure shares and saves, then brand searches a week later. Views tell you nothing.

The limits

• Rights. The image in a meme often belongs to a photographer or a studio, and commercial use is not private use. On top of that come people's right to their own image (Art. 28 of the Swiss Civil Code) and music, which should be taken from the catalogues reserved for business accounts.
• AI. A generated video that resembles a real person or a real event must be labelled (section 30).
• Tone. Insider humour works in B2B, but never at customers' expense.
• Audience. If your buyers are not on these platforms, a view is worth nothing.

Applied to LEEZ

What it brings. Saying that you are selling your business is still taboo, and many people in their thirties think that being an entrepreneur means starting from scratch. A light tone can get two ideas across: selling is not a failure, and taking over a business is one way into entrepreneurship.

In practice. SME owners and fiduciaries (Swiss accounting and trust firms) are found on LinkedIn more than on TikTok, even though Instagram remains the leading network in Switzerland according to Digimonitor 2026. A business transfer touches on the end of a career: the right tone is tenderness, not mockery. Two formats to test: a short, anonymised series, “a business to take over in sixty seconds”, and buyers' testimonials in the spirit of Tayo's web series “Moteur!” (section 6).

A meme is first of all an image. And an image published in 2026 is read, described and sometimes reused by AI systems: should you let it be found?

↑ Contents

Section 28

Images in the age of AI

How an image gets found

Google reads the page before the image. Its documentation asks you to embed images with the standard HTML tag, and states that it infers their subject from the content of the page: nearby text, caption, title. The same image can then appear in the Images tab, as a thumbnail next to a link, in Discover, in Google Lens and in generated answers.

Visual search has changed scale. Google reports more than 25 billion searches a month in Lens in March 2026, one in five of them with commercial intent. These are its own figures, unaudited. The phone camera has become a search bar: a photo of a product, a dish or a place is the way in.

AI assistants read the pixels: objects, embedded text, charts. But to choose and cite an image, they rely mainly on what surrounds it: the page, the caption, the alt text, the metadata. No independent study yet measures the weight of each of these signals.

The bots that come for images

Two traps. Googlebot serves both Google's search and its generated answers: you cannot refuse the second without losing the first. And OpenAI states that ChatGPT-User, which is triggered by a user, might not apply the robots.txt file.

Getting an image ranked, and picked up by AI assistants

1. An original, useful image: my photo, my diagram, my chart. A stock image found on ten thousand sites has no reason to be chosen.
2. Placed near the text it illustrates, with a visible caption.
3. A real image tag, with alt text that describes what you see. Never a background image.
4. A descriptive file name: atelier-menuiserie-vaud.webp rather than IMG_7489.jpg.
5. A lightweight format, WebP or AVIF, served in several sizes.
6. No lazy loading on the main image, which weighs on rendering time (section 18).
7. A stable, unique URL per image, listed in an image sitemap.
8. Rights metadata: author, credit and licence, in structured data and in the file itself. Google can then display the “Licensable” badge.
9. A share image declared in the page, at least 1,200 pixels wide for Discover.
10. A robots.txt file that lets through the search crawlers whose visits I want.

An assistant only cites what its search crawler can read. Blocking GPTBot does not stop you being cited in ChatGPT; blocking OAI-SearchBot does.

My observation, which no one has yet put a number on: charts of original data, annotated diagrams and tables are picked up more often than illustrative photos, because they carry information that can be verified.

Saying where an image comes from

In Switzerland, no law currently requires you to label a generated image. The Federal Council is preparing a draft AI regulation for the end of 2026. A Swiss company that distributes content to the European Union is still covered by Article 50 (section 30).

Protecting your images: three goals not to confuse

1. Not being used for training: keeping the work out of a dataset.
2. Not being reused in answers: stopping an assistant from showing the image in place of the site.
3. Not being imitated: stopping anyone from generating “in the style of”.

A method that works against one may be useless against another.

Glaze and Nightshade, two tools from the University of Chicago, add an invisible perturbation to the image: the first disrupts the imitation of a style, the second poisons training. They have been downloaded several million times. Their limit is documented: in 2025, researchers showed that simple methods get round them, and the LightShed tool detects Nightshade in 99.98% of cases before removing the perturbation.

Since 15 September 2026, Cloudflare has blocked training crawlers and agents by default on advertising-funded pages, for new sites and free accounts. The setting needs care: for a dual-use bot such as Googlebot, the strictest rule applies, and blocking training can block search.

What the law says

• In Switzerland, the Copyright Act (CopA) provides a text and data mining exception only for scientific research (Art. 24d CopA). Commercial training is not expressly covered. Parliament has adopted the Gössi motion, which calls for better protection of intellectual property in the face of AI: a bill is expected.
• In the European Union, the 2019 copyright directive allows text and data mining, unless the author has expressed a machine-readable reservation. In Hamburg, in Kneschke v LAION, the court of appeal ruled in December 2025 that a reservation written in natural language was not enough for the period in question.
• In Munich and London, the two November 2025 judgments presented in section 20 draw the line: outputting a work memorised by the model is a reproduction (GEMA v OpenAI), but a model's weights are not a copy (Getty Images v Stability AI).
• For individuals, Art. 28 of the Swiss Civil Code protects a person's image and voice: the victim of a deepfake can demand that it stop and claim redress.

My view

Real protection comes down to three moves: not publishing in high resolution what you do not want to see copied, setting a machine-readable reservation that will carry weight in court, and sorting the bots one by one. The rest is mostly symbolic: “noai” tags, opt-out registries whose observance no one measures, and the idea that a filter makes a work impossible to learn.

The trade-off is simple to state and hard to live with. The artist who lives on commissions has to be found: they block training, leave search open, watermark their previews and keep their originals offline. The artist whose style is the product protects themselves more and accepts being less visible.

Applied to LEEZ

What it brings. An owner only puts their business up for sale if they are sure of staying anonymous. Yet a single photo can give them away: a reverse image search finds a shopfront in seconds, and the GPS coordinates lie dormant in the file. Handled well, images still serve business transfers: the barometer charts and the maps by canton can be picked up by Google and by AI assistants, and get businesses for sale noticed.

In practice.

1. Strip location metadata on upload, server-side, without relying on the seller.
2. Refuse building fronts, signs, shop windows, sign-written vehicles, rare machines and views through the window.
3. No recognisable person, neither employee nor customer.
4. Prefer illustrative visuals, flagged under the image: “Illustrative image, does not show the business”.
5. Run a reverse search before every publication: if the image leads to the business's website or its Google Maps page, it is refused.
6. Never reuse a photo already online on the seller's side, even cropped.
7. Close listing images to indexing and to AI crawlers, and leave the listing text, the barometer and the guides open.

With your images sorted, you still have to get talked about. A press article now has one more reader: the AI that will cite it.

↑ Contents

Section 29

Media coverage: is buzz still worth anything?

The benefits in 2026

• AI assistants cite what you do not own. Muck Rack analysed more than 25 million links cited by ChatGPT, Claude and Gemini (May 2026): 84% come from sources that brands neither own nor pay for, including 27% from journalism. Brand websites account for 13.7%, paid content for 0.3%.
• A press release on its own is useless. According to BuzzStream, press releases distributed through newswires account for 0.04% of citations.
• Journalists search with AI. 82% of them use it for their research, according to Muck Rack. Being cited by an AI also means being found by the next journalist.
• The proof can be reused on the website (section 12), in partnership proposals and with banks.
• People search for the brand. A mention lifts brand searches and direct traffic, the two doors that are holding up (section 11).

Two caveats. Muck Rack sells PR software, and points out itself that this figure does not mean PR accounts for 84% of visibility: journalism's share is 27%. What is more, only 2% of the journalists that PR people usually pitch are among those AI assistants cite. Aim for specialist and regional titles, not the big names.

Is buzz picked up by the media a good thing?

Two facts weigh in the balance. The sources AI assistants cite turn over every one to three months (section 15): a one-off fades, a steady rhythm lasts. And digital memory is permanent: buzz that went wrong resurfaces every time someone checks up on you, which is costly in a market as small as French-speaking Switzerland.

Should you create campaigns nobody has seen before?

Yes for the angle, no for the stunt in itself. I distinguish three levels.

1. New data, the most profitable. A barometer, a study, a ranking that no one else can publish. It is repeatable, cited for a long time, and risk-free.
2. A new format. A new way of showing something true: the Burgerpass as an object (section 4), a work of art drawn from Tayo's source code (section 6). New here is enough: a format proven elsewhere is a novelty for a journalist in French-speaking Switzerland.
3. The pure stunt. Very hit-and-miss. Only attempt it if it is true to the brand, with a crisis plan ready.

Before launching, five questions:

• Can the story be told without us?
• Which figure will still be cited in six months?
• Who could it harm?
• What do we do if it takes off too much, or not at all?
• How do we recycle it: radio, podcast, a page on the site (section 25)?

Applied to LEEZ

What it brings. The story of a baker who found his successor convinces ten other owners that it can be done. And a figure published every quarter puts business transfers into the public debate, which is where support schemes, guarantees and training for buyers are decided.

In practice.

• The recurring campaign: the quarterly barometer (section 15), sent first to the regional and specialist press.
• The stories: a seller and their buyer, with their consent, for the daily papers and radio stations of the cantons.
• A new format: a map of businesses for sale by canton and by sector, updated every quarter. It respects the anonymity of the listings, and it can be shared and cited.

You still need someone to carry the story. Some brands have chosen to manufacture that face rather than find it.

↑ Contents

Section 30

AI-generated ambassadors

Practices tested in several industries

Fashion: H&M (2025). H&M worked with the Swedish company Uncut to create “digital twins” of 30 real models, including Mathilda Gvarliani and Vilma Sjöberg. The models keep ownership of their twin, can license it to other brands and are paid per use; the images carry a watermark. The first campaign appeared on 2 July 2025. The British union Bectu and model advocates denounced the risk to jobs, particularly in e-commerce. No commercial results with figures have been published.

Fashion: Guess in Vogue (summer 2025). A two-page advertisement for Guess in the August 2025 issue of US Vogue used models generated by the London agency Seraphinne Vallora, with no disclosure other than a small line indicating AI production. The model Felicity Hayward described the move to the BBC as “lazy and cheap”; the British eating disorder charity Beat called it “worrying”, and readers said they were cancelling their subscriptions. Condé Nast stated that no AI model had ever appeared in Vogue's editorial content.

Influencer marketing: Aitana López (The Clueless, Barcelona). According to her creators, this virtual influencer can earn up to €10,000 a month, with an average of around €3,000 (Euronews, March 2024). These figures are self-reported and date from 2023–2024.

Health and causes: Lil Miquela and NMDP (2025). The virtual influencer Lil Miquela ran a storytelling campaign with NMDP (the US bone marrow donor registry) in which she is diagnosed with leukaemia; according to MM+M, the campaign generated more than 5 million organic impressions. The figures come from the organisation, and the narrative choice is debated.

Film: Tilly Norwood (2025). The AI “actress” created by Particle6/Xicoia caused an outcry in Hollywood in late September 2025. The union SAG-AFTRA stated that Tilly Norwood is not an actor, but a character generated by a computer program trained on the work of countless professional performers, without permission or compensation.

Tourism: “Emma”, German National Tourist Board (October 2024). An interactive AI avatar presented as a brand ambassador, developed for around €60,000 according to a trade website; its Instagram account had barely more than 1,000 followers at launch, and travel creators protested. The AI influencer “Venus” from the Italian campaign “Open to Meraviglia” is also reported to have generated limited engagement.

Banking and automotive. All you find are vendor case studies (avatars in bank branches, for example) with no independently verified results. Treat them as sales pitches.

What the research says

• The “Muse Two” study by Billion Dollar Boy (fieldwork by Censuswide, June 2025, 4,000 consumers in the United States and the United Kingdom) indicates that 76% of consumers trust virtual influencers for product recommendations, but that only 26% prefer AI-generated creator content (against 60% in 2023) and that 57% think digital twins erode trust. Contradictory results, from an agency with a commercial interest.
• A study published in the Journal of Business Research (2025, 255 participants) shows that when a virtual influencer promotes a faulty product, the public blames the brand rather than the influencer: AI influencer marketing can do more damage to trust than human influencer marketing.

Transparency obligations

• AI Act, Article 50: applicable since 2 August 2026. AI systems that interact with people must disclose that they are AI; deepfakes (generated or manipulated images, audio or video that resemble real people, places or events) must be disclosed as such. The “Digital Omnibus” Regulation (EU) 2026/1744, which entered into force on 27 July 2026, postponed the obligations for high-risk systems (Annex III) to 2 December 2027, but did not postpone Article 50; only the machine-readable marking obligation (Art. 50(2)) has an extension until 2 December 2026 for systems already on the market. Penalties can reach €15 million or 3% of global turnover.
• Platforms: Meta, TikTok and YouTube require realistic AI-generated content to be labelled.
• Switzerland: no equivalent specific law to date, but the Unfair Competition Act (UCA) penalises misleading statements, the Federal Act on Data Protection (FADP) protects image and voice as personal data, and personality rights (Art. 28 of the Swiss Civil Code) protect against unauthorised use of a person's image or voice. A Swiss company that targets the EU market is covered by the AI Act.

My view

AI ambassadors work for high-volume e-commerce (variants of product visuals), provided the people cloned consent and are paid. They fail when they replace a human relationship without saying so. A simple rule: if you would be ashamed for the public to find out, disclose it or do not do it.

Applied to LEEZ

What it brings. Almost nothing, and that is the point of the example. The sale of a life's work is entrusted to people. A synthetic face would weaken the trust on which every business transfer depends.

In practice. No synthetic ambassador to carry the brand. Two uses remain acceptable, provided they are disclosed: explainer videos produced in four languages with a synthetic voice, and illustrative visuals for anonymous listings, never presented as photos of the business.

The synthetic face is only the most visible temptation. Another is more discreet and far more serious: using data that is lying around.

↑ Contents

Section 31

Data leaks and building a “Cambridge Analytica of your own”: the real risk!

This section deals with the risk. It gives no instructions for acquiring or exploiting stolen data.

The temptation

An experienced growth hacker, equipped with AI agents, could be tempted by a shortcut: cross-referencing leaked databases, enriching profiles, inferring psychological traits and micro-targeting messages. Technically, AI lowers the barrier: text classification, attribute inference and the generation of personalised messages at scale have become trivial. That is precisely why it needs to be talked about.

What is circulating

Massive leaks keep coming: compilations of billions of credentials circulate regularly, often aggregated from older leaks and from credential-stealing malware (“infostealers”). Switzerland is not spared: in 2023, the hack of the service provider Xplain led to federal administration data being published on the darknet. I give no overall figure for 2025–2026 here, because the counts published by cybersecurity companies often mix duplicates and old leaks.

The Cambridge Analytica reminder

In 2018, it emerged that Cambridge Analytica had exploited the data of up to 87 million Facebook users, collected through a quiz app, for political targeting. The consequences: the company's bankruptcy, a $5 billion fine imposed on Facebook by the FTC in 2019, and a £500,000 fine from the UK's ICO.

What the research says about effectiveness. A study by Matz et al. (PNAS, 2017) showed that ads tailored to inferred personality could increase clicks and purchases in controlled experiments. But many researchers have pointed out that Cambridge Analytica's real electoral effectiveness was greatly overestimated, partly by the company itself for commercial purposes. The lesson: the danger lies less in all-powerful manipulation than in the mass violation of rights and the loss of trust.

The legal and criminal framework

• Federal Act on Data Protection (FADP, Switzerland): unlawful data processing, breach of the duties to provide information; criminal fines of up to 250,000 Swiss francs, aimed at the individuals responsible.
• Swiss Criminal Code (SCC): unauthorised access to a data processing system (Art. 143bis SCC), unauthorised obtaining of data (Art. 143 SCC) and the use of unlawfully obtained data can be prosecuted; whether handling stolen goods applies to data is debated among legal scholars, but using stolen data exposes you to prosecution on several grounds.
• GDPR (EU): fines of up to €20 million or 4% of global turnover; processing of sensitive data (political opinions, health) subject to strict conditions; profiling regulated.
• DSA (EU): bans online platforms from running targeted advertising based on sensitive data, and targeted advertising aimed at minors.

Reputational risks

A single revelation is enough to destroy a brand. In a market as small as French-speaking Switzerland, reputation is an asset that cannot be replaced.

The legal (and more effective) alternatives

• First-party data: behaviour on your site and in your product, with consent.
• Zero-party data: what your customers tell you voluntarily (quizzes, preferences, surveys).
• Lawful enrichment: compliant B2B providers, public registers (Zefix), company data.
• OSINT (open-source intelligence) with safeguards: public sources, legitimate purpose, proportionality, documentation.
• Synthetic data: to test models without exposing real people.

My position: no growth hack is profitable enough to justify a criminal record and a destroyed brand. Powerful tools demand stronger ethics, not weaker ones.

Applied to LEEZ

What it brings. An owner can only sell with peace of mind if no one finds out too early. If their name leaks, their employees, customers and competitors know they are leaving, and the business loses value before it is even sold. Confidentiality is not a technical constraint: it is the precondition for the goal.

In practice. A business transfer platform is a target before it is a user of data.

• Collect the minimum, and partition the data room deal by deal.
• Keep an access log and have the platform tested every year (section 12).
• Prepare the notification: who to notify, how quickly, with what message.
• Buy no list of company directors, whatever its stated source.

Refusing stolen data does not settle the question of persuasion. To know where the line runs, you have to look at those who crossed it furthest.

↑ Contents

Section 32

What the propaganda of authoritarian regimes teaches (and what you must never copy)

A deliberately provocative section, but a rigorous and critical one. It is in no way an apologia for these regimes, which are responsible for mass crimes: the aim is to understand mechanisms of persuasion in order to protect yourself against them and to tell what is legitimate from what is not.

The reference literature

• Edward Bernays, Propaganda (1928): Freud's nephew founds modern public relations, then in 1947 theorises the “engineering of consent”. His work was read by Goebbels, a reminder that the tools of persuasion are neutral and their uses are not.
• Jacques Ellul, Propagandes (1962): modern propaganda is total and continuous, and answers individuals' need to make sense of things in a mass society.
• Hannah Arendt, The Origins of Totalitarianism (1951): totalitarian propaganda aims less to convince than to create a coherent fictional world in which facts no longer matter.
• Edward Herman and Noam Chomsky, Manufacturing Consent (1988): even in democracies, structural “filters” shape information.
• RAND, “The Russian ‘Firehose of Falsehood’ Propaganda Model” (Paul and Matthews, 2016): high volume, multiple channels, speed, repetition, no commitment to truth or consistency.
• Oxford Internet Institute, work on computational propaganda: documents the organised use of automated accounts, trolls and astroturfing campaigns by governments and parties in many countries.

The mechanisms, analysed

The lessons worth keeping

1. Clarity wins: a simple, repeated, consistent message beats a brilliant but shifting one.
2. Rituals create belonging: a regular fixture is worth more than a one-off campaign.
3. Story comes before argument: people remember stories, not feature lists.
4. Saturation without truth ends up destroying trust: regimes that lie lose their credibility when reality catches up with them, and so do brands, only faster.

The red lines

• No factual lie, no fake review, no fake account (Swiss UCA, EU Unfair Commercial Practices Directive).
• No naming of a group of people as the enemy.
• No targeting based on sensitive data without a legal basis (GDPR, Swiss Federal Act on Data Protection, DSA).
• No undisclosed AI ambassador or bot (AI Act, Article 50).

My conclusion: you can borrow propaganda's techniques of form (clarity, repetition, rituals, storytelling) provided you put them at the service of the truth and of the customer's freedom of choice. The day a tactic only works if the public does not know what you are doing, it is manipulation.

Applied to LEEZ

What it brings. For an owner to think of passing the business on rather than closing it, they need to have heard, several times and in simple terms, that it is possible. The honest repetition of a clear message is enough for that.

In practice.

• One message, repeated everywhere: pass on your business without being exposed.
• An opponent that is a problem, not a group: healthy businesses closing for want of a buyer.
• A ritual: the barometer, on a fixed date, every quarter.
• The red line specific to marketplaces: never inflate supply or demand. No fictitious listing, no fake buyer, no doctored counter. It is the classic temptation of the early days, and it is deception within the meaning of the UCA. A buyer who has been deceived once does not come back.

The red lines are drawn. What remains is open ground: here is what I see least exploited around me.

↑ Contents

Section 33

Blind spots and unexplored avenues

1. Optimising for shopping agents. Make your catalogue readable and purchasable by agents (complete schema.org Product markup, UCP via Shopify, ACP via Stripe). Almost no one in French-speaking Switzerland does it yet.
2. Distribution through MCP directories and assistant apps. Listing an MCP server in the official registry and offering an app in ChatGPT or Claude means occupying a new store before the crowd arrives.
3. Proprietary data as a citation asset. An annual barometer, a price index, a survey: the currency of GEO.
4. Non-English-language markets. Most GEO tools and content are designed in English. Swiss French, with its legal and cultural specifics, is a citation market with little competition. A word of caution: according to the panel cited in section 15, a French prompt triggers searches in French and in English, so you need to exist in both languages.
5. Physical and local channels. Regional radio on the rise, handwritten mail, events: their relative costs are falling against saturated digital channels.
6. Free tools as magnets. A calculator, a generator, a free audit, all the easier to build with vibe coding, and naturally cited by AI assistants.
7. Skills and MCP servers as lead magnets. Publish a free Agent Skill (for example “GEO audit using method X”) that makes your method known every time it runs. That is exactly what ChatSEO does by publishing its GEO skills.
8. Private communities on WhatsApp and Telegram. Messaging apps have become the very interface of agents (OpenClaw is controlled via WhatsApp and Telegram): channels and communities there take on new value.
9. Programmatic SEO revisited. Pages generated at scale but fed with real, unique data (like the dog park directory demonstrated by Nicholas Dulait), not empty text.
10. Voice. Transcribed podcasts, radio segments, the voice interfaces of assistants: an expert's voice is a signal of authenticity that is hard to imitate credibly over the long term.
11. Crawler control as a bargaining lever. With pay-per-crawl or pay-per-use models, a niche publisher could monetise access to its quality content.
12. Compliance as a selling point. Being visibly compliant with Article 50 of the AI Act and with the Federal Act on Data Protection (FADP) reassures business customers, who have to comply themselves.
13. Fan-out as keyword research. Track the short queries that AI assistants type into Google rather than users' prompts. The method has only recently become public and almost no one applies it in French (section 15).
14. Small niche sites rather than major media. According to the panel in section 15, a mention on a small, relevant site is cited for as long as a mention that costs around a hundred times more in a major outlet. A network of niche partners, open and transparent, is worth more than a press coup.

Applied to LEEZ

What it brings. These avenues have one thing in common: they put a business for sale in front of people who were not yet looking for it.

In practice. Three of them come first: the data that only LEEZ holds (the barometer), the short queries that AI assistants generate in place of its customers, and a presence in connector directories, where a buyer will soon ask their assistant to search on their behalf.

Three levers remain, older and quicker to launch. The first is as old as the web: the forum.

↑ Contents

Section 34

The old-school forum: a perfect hack?

Why forums matter again

• AI assistants read them. According to an OtterlyAI analysis of more than a million citations, community forums account for between 5.9% and 16.9% of cited sources, depending on the assistant.
• The format suits them. A precise question followed by an answer drawn from experience is exactly what a model knows how to extract.
• What is said on Discord or WhatsApp is invisible to search engines and to AI assistants. A public forum stays indexable for years.
• Experience cannot be made up. An AI cannot tell you what the bank asked of the person who took over a bakery in Morges.

One caveat: the figures on Reddit contradict each other. Tinuiti measured a sharp rise in Reddit citations in late 2025; the panel in section 15 measured a drop in ChatGPT's French-language answers in summer 2026. My reading stays the same: a big platform saturated by marketers ends up downgraded, while a small, precise niche forum keeps being cited.

The hack: answer where the question has already been asked

I found no Swiss forum devoted to business transfers. The discussions exist, but elsewhere.

Also worth exploring: forums for cross-border workers, German-speaking Swiss forums for entrepreneurs and, above all, trade forums. With 1,465 hotel and restaurant listings, the places where restaurant owners talk to each other are worth more than any general forum. Before writing there, read each forum's rules on company participation.

The method comes down to five moves:

1. List thirty existing threads, old ones included, that ask a question LEEZ has data on. An old thread that ranks well is still read, by humans and AI assistants alike.
2. Answer under a real name and job title, with a figure from the barometer, and with no link in the answer. The link lives in the profile signature.
3. Publish the long answer on leez.ch, and point to it only if someone asks for it or the forum rules allow it.
4. Offer an open Q&A session, announced to the moderators, on a specific subject: for example, taking over a restaurant in Geneva.
5. Measure brand searches, visits from the forum and AI citations on the queries concerned.

The red lines are those of section 32: no fake accounts, no fake testimonials, and any link with LEEZ always disclosed.

Should LEEZ open a forum of its own?

My recommendation: “Business Transfer Questions”, not an open forum

• Anonymous questions, asked under a pseudonym, with the canton and the sector as the only context, and reviewed before publication.
• Signed answers from partner experts, fiduciaries (Swiss accounting and trust firms), lawyers and business brokers, with a link to their profile.
• One page per question, marked up as a Q&A page.
• Fifty real questions at launch, drawn from support and from meetings, so that the space is never empty.
• French first, German next.

A stopping rule, set in advance: if the space does not receive twenty or so new questions a month after six months, LEEZ closes question submissions and keeps the archive as an FAQ.

The safeguards: a notice reminding readers that the answer is no substitute for personalised advice, no company names, and moderation before publication. Authors' data is processed in accordance with the Federal Act on Data Protection (FADP).

The forum is not the only old idea finding a new use. A whole swathe of the pre-2015 web is waiting to be brought back out, and a console from 1989 explains why.

↑ Contents

Section 35

Forgotten technologies (1990–2015): what Nintendo teaches us

The lesson of the Game Boy

In 1989, Nintendo launched a handheld console that looked outdated in every respect. The Game Boy's screen was monochrome, greenish and unlit. Its processor was an 8-bit chip from an already old generation. Its rivals, Atari's Lynx and then Sega's Game Gear, displayed colours on a bright screen.

The choice was deliberate. It bears the signature of Gunpei Yokoi, who joined Nintendo in 1965 to maintain the assembly-line machines, became a toy inventor, then the father of the Game & Watch in 1980: pocket games built with calculator screens, which by then were cheap and thoroughly mastered. Yokoi gave his method a name, usually translated as “lateral thinking with withered technology”. The principle: do not chase cutting-edge technology, but take a mature one, which is therefore cheap, reliable and understood by everyone, and find a use for it that nobody had thought of.

Turning down the colour screen, which was the cutting-edge technology of the day, gave the Game Boy its battery life and its price. The public did not want the best-looking screen: it wanted to play for a long time, anywhere, without breaking the bank. The Game Boy and its colour successor sold 118.69 million units. The Game & Watch had already sold 43.4 million. The rivals disappeared.

I take three ideas from this for growth:

• Mature technology is cheap, reliable and compatible with everything. Nobody is looking at it, so nobody competes with you there.
• The advantage does not come from the tool, but from the use you invent for it. That is the “lateral” half of the formula, and the harder one.
• Timing matters. An old technology becomes valuable again when the world around it has changed. In 2026, the simple formats of the old web happen to be the ones AI agents read best.

Why the old web is useful again

• Most AI crawlers do not execute JavaScript (section 17). Pre-2010 formats are readable by design.
• The audience is at home with it. A seller reads emails, prints things out and picks up the phone.

To bring back right away

To adapt

To leave dormant

• The open forum: see section 34.
• Webrings: chains of reciprocal links, which Google now counts among link schemes.
• OpenSearch (2005): the ancestor of opening a site up to machines, replaced by MCP and WebMCP (section 14).
• AMP (2015): the lesson stays, lightweight pages in simple HTML; the format does not.
• Flash, toolbars and pop-ups: no.

Where to start

1. The feed per canton and per sector. It is the basic building block of the alerts and of a future MCP server (section 14).
2. The plain-text email alert.
3. The glossary in four languages.

These technologies proved themselves twenty years ago. And today, for those starting from zero? 290 founders have published their numbers.

↑ Contents

Section 36

More ideas, drawn from Indie Hackers

What 290 founders say

In late September 2026, Isaac Harold went through a year of founders' posts about their revenue in twelve Reddit communities: 290 founders, 209 of whom say where their first paying customers came from.

Three lessons:

• Direct outreach opens the door, it does not drive growth. Past $10,000, content and SEO come up most often as the growth engine. Direct outreach is the main channel for only 16% of them.
• The targets are narrow. Those who succeed through direct outreach go after office managers at dental practices or wedding photographers, never “small businesses”.
• Paid advertising is the most cited failure: 87 of the 144 founders who name a failure mention it.

The author flags the biases himself: unverified claims, Reddit over-represented, and only those who succeeded post. For consumer products, the result is reversed: none started with direct outreach.

Ten ideas and their examples

These revenue figures are reported by the founders themselves, and some of the articles are for subscribers only: I cite them as testimonies, not as verified data.

Applied to LEEZ

What it brings. These founders show that a small team reaches a lot of people when others do the distribution for it. For business transfers, that means going through the people the owner already listens to.

In practice.

• Fiduciaries (Swiss accounting and trust firms) are LEEZ's referrers. Like Zigpoll's agencies, they are the ones who bring it to their clients. Nothing should cost them anything.
• A narrow target for the next sellers. Geneva restaurant owners nearing retirement rather than “SME owners”, and by handwritten letter (section 24) rather than by mass email, which the Unfair Competition Act (UCA) regulates.
• Two questions to ask. At sign-up: how did you hear about LEEZ? After the valuation: what nearly made you give up?
• Acquisition as a lever. Taking over a directory, a newsletter or a small listings site can cost less than building their audience. It is LEEZ's own trade, applied to itself.

All these levers lead to the same question: if orchestrating agents is enough, what is left of the job?

↑ Contents

Section 37

Conclusion: the future of the growth hacker's job in 2026

From “hacker” to orchestrator

The growth hacker of 2015 looked for the loophole in a channel. The growth hacker of 2026 orchestrates agents that execute, test and measure, while concentrating on three things: the choice of bets, the quality of judgement and trust in the brand.

Skills on the rise

• Agent orchestration: MCP, skills, automations, guardrails.
• Data: SQL, statistics, measuring AI visibility.
• Editorial: the ability to produce original content and data that machines cite.
• Human: community building, negotiating partnerships, being there in person.
• Legal: the Federal Act on Data Protection (FADP), the GDPR, the AI Act and the Unfair Competition Act (UCA) as a reflex, not as a brake.

Skills in decline

Generic content production, manual campaign management, repetitive reporting: agents already do them.

The job market

The title “growth hacker” is becoming rare; its duties are being shared out among growth engineers, lifecycle marketers, GEO specialists and heads of marketing AI. Teams are shrinking (one person and ten agents), but demand stays strong for people who can connect product, data and narrative. This reading is an analysis, not a statistic.

What remains human

Taste, the courage to say no, relationships of trust, meeting in person, ethics. AI assistants can generate a thousand ideas; they do not know which one is worth risking your reputation for.

Outlook 2027–2030

• 2027: shopping agents become a measurable channel in e-commerce; the standards (MCP, UCP, ACP, WebMCP) consolidate.
• 2028: visibility in assistants weighs as much as classic SEO in acquisition budgets.
• 2029–2030: value concentrates in three assets: proprietary data, community and trust. The growth hacker becomes an architect of these three assets.

My conviction: those who only knew how to execute will be replaced by agents; those who know how to decide, tell a story and bring people together will be more sought after than ever.

Applied to LEEZ

What it brings. Machines find, sort and connect. The decision to hand over the business of a lifetime, though, is made between people. The whole job is to leave to machines whatever saves time, and to give that time back to meeting people.

In practice. A small team, agents for execution (the site audit, citation monitoring, the monthly check), and people for what cannot be delegated: convincing an owner to entrust the sale of their business, or a trade association to recommend LEEZ.

Let's continue on LinkedIn

This guide will evolve with what the LEEZ case teaches me. To follow what comes next, contradict me or tell me what works for you, join me on LinkedIn: linkedin.com/in/johanbavaud.

↑ Contents

Sources

The sources are named in the text of each section. Here are the main web addresses.

• Nicholas Dulait on X, Links Garden panel data and GEO method: first post (https://x.com/NicholasDulait/status/2105931560394330199), second post (https://x.com/NicholasDulait/status/2093265964552856015)
• Billion Dollar Boy, “Muse Two” study on AI and the creator economy: billiondollarboy.com (https://www.billiondollarboy.com/news/new-research-real-impact-ai-creator-economy)
• Hive Security, the OpenClaw security crisis: hivesecurity.gitlab.io (https://hivesecurity.gitlab.io/blog/openclaw-ai-agent-security-crisis-2026/)
• Ivinco, comparison of the UCP, ACP and MCP protocols: ivinco.com (https://www.ivinco.com/blog/protocol-wars-ucp-acp-mcp)
• One FM and Radio Lac, audience figures and DAB+: onefm.ch (https://www.onefm.ch/categories/dab/feed/), radiolac.ch (https://www.radiolac.ch/categories/dab/feed/)
• MCP, specification of 28 July 2026: blog.modelcontextprotocol.io (https://blog.modelcontextprotocol.io/posts/2026-07-28/)
• WebMCP, Chrome documentation: overview (https://developer.chrome.com/docs/ai/webmcp), declarative API (https://developer.chrome.com/docs/ai/webmcp/declarative-api)
• ACP: specification on GitHub (https://github.com/agentic-commerce-protocol/agentic-commerce-protocol), OpenAI documentation (https://developers.openai.com/commerce), Digital Commerce 360 on the change of course in March 2026 (https://www.digitalcommerce360.com/2026/03/06/openai-shifts-checkout-plans-agentic-commerce-strategy/)
• UCP: ucp.dev (https://ucp.dev/), core concepts (https://ucp.dev/documentation/core-concepts/), expansion of the technical council (https://ppc.land/amazon-meta-microsoft-salesforce-and-stripe-join-ucp-tech-council/), March 2026 update (https://www.semrush.com/blog/universal-commerce-protocol/), May 2026 announcements (https://searchengineland.com/google-expands-universal-commerce-protocol-and-launches-new-agentic-shopping-tools-478113)
• The LEEZ case: leez.ch, swisspeers lists (https://info.swisspeers.ch/firmensuche-schweiz)
• Embed and SEO: Google's spam policies (https://developers.google.com/search/docs/essentials/spam-policies), JavaScript and SEO (https://developers.google.com/search/docs/crawling-indexing/javascript/javascript-seo-basics), qualifying outbound links (https://developers.google.com/search/docs/crawling-indexing/qualify-outbound-links), BizBuySell, listings on the business broker's website (https://www.bizbuysell.com/blog/how-to-make-your-bizbuysell-listings-appear-on-your-website/)
• AI crawlers and JavaScript: summary by seobro.com (https://seobro.com/blog/do-ai-crawlers-execute-javascript/)
• Fashion ID judgment: press release from the Court of Justice of the EU (https://curia.europa.eu/site/upload/docs/application/pdf/2019-07/cp190099en.pdf)
• Devices: StatCounter, September 2026, worldwide (https://gs.statcounter.com/platform-market-share/desktop-mobile-tablet) and Switzerland (https://gs.statcounter.com/platform-market-share/desktop-mobile-tablet/switzerland)
• Zero-click searches: SparkToro and Similarweb study (https://sparktoro.com/blog/in-2026-less-than-one-third-of-google-searches-still-send-a-click/)
• Traffic sent by AI assistants: SE Ranking (https://seranking.com/blog/chatgpt-referral-traffic-may-2026/), summary of the Conductor study (https://www.tryanalyze.ai/blog/ai-traffic-research), conversion measured by Adobe (https://www.elmohq.com/blog/ai-referral-traffic-conversion)
• Bots and humans: Cloudflare Radar data from June 2026 (https://www.digitalapplied.com/blog/ai-crawler-bot-traffic-statistics-2026-data-reference)
• Usage in Switzerland: IGEM Digimonitor 2026 (https://www.igem.ch/digimonitor/)
• Forums and AI citations: summary of the OtterlyAI figures (https://www.getpassionfruit.com/blog/how-ai-search-treats-user-generated-content-reviews-forums-and-community-posts), Tinuiti report on Reddit (https://www.cmswire.com/digital-marketing/reddits-rise-in-ai-citations-what-marketers-must-know-about-aeo-strategy/)
• Swiss forums: English Forum Switzerland (https://www.englishforum.ch/business-entrepreneur/), Mustachian Post (https://forum.mustachianpost.com/t/has-anyone-bought-a-small-business-in-switzerland/12605)
• Site audit: Chrome DevTools MCP server (https://github.com/ChromeDevTools/chrome-devtools-mcp), Core Web Vitals thresholds (https://yassersoliman.com/blog/core-web-vitals-explained-marketers/), the disavow tool in 2026 (https://almcorp.com/blog/google-disavow-tool/)
• Indie Hackers: the analysis of 290 founders (https://www.indiehackers.com/post/i-went-through-290-founders-mrr-posts-the-ones-past-10k-got-their-first-customers-a-different-way-ff0f3e5904), the Zigpoll case (https://www.indiehackers.com/post/tech/hitting-125k-mrr-as-a-solo-founder-by-doubling-down-on-the-right-segment-c4o2Tfs6mjdpip5yZhaO), the other examples (https://www.indiehackers.com/)
• Credibility: TrustedSite 2026 survey (https://www.trustedsite.com/resources/trust-badges), Edelman Trust Barometer 2026, report on brands (https://www.edelman.com/trust/2026/trust-barometer/special-report-brands)
• Press and AI citations: Muck Rack study presented by the PRSA (https://prsay.prsa.org/2026/10/01/ai-citations-spark-renaissance-moment-for-pr-and-communications), clarification of what the 84% figure measures (https://machinerelations.ai/research/why-ai-citation-studies-disagree-source-register-2026), summary including the BuzzStream figure (https://www.shadow.inc/resources/earned-media-ai-citations)
• Viral campaigns: fifteen cases analysed by DesignRush (https://www.designrush.com/agency/digital-marketing/trends/viral-marketing-campaigns)
• LEEZ: about page
• Experimentation: success rates compiled by Ronny Kohavi (https://www.growthbook.io/blog/designing-a-b-testing-experiments-for-long-term-growth)
• True fans: Kevin Kelly's essay (https://kk.org/thetechnium/1000-true-fans/), Li Jin's extension (https://a16z.com/1000-true-fans-try-100/)
• Swiss SMEs: figures from the Swiss Confederation's SME Portal (https://www.kmu.admin.ch/kmu/fr/home/savoir-pratique/politique-pme-faits-et-chiffres/chiffres-sur-les-pme/entreprises-et-emplois.html)
• Game Boy: Gunpei Yokoi (https://en.wikipedia.org/wiki/Gunpei_Yokoi), Game Boy (https://en.wikipedia.org/wiki/Game_Boy), Game & Watch (https://en.wikipedia.org/wiki/Game_%26_Watch)
• Images and AI: Google's best practices for images (https://developers.google.com/search/docs/appearance/google-images), OpenAI's crawlers (https://developers.openai.com/api/docs/bots), Cloudflare's new policy (https://techcrunch.com/2026/07/01/cloudflares-new-policy-pushes-ai-companies-to-pay-for-publishers-content/), limits of Glaze and Nightshade (https://poisoning.ai/articles/lightshed-explained/)
• Images and the law: AI regulation in Switzerland, update from SUISA (https://blog.suisa.ch/fr/ki-regulierung-stand-der-dinge/), GEMA v OpenAI (https://se-legal.de/gema-openai-copyright-case-germany/?lang=en), Getty Images v Stability AI (https://www.mishcon.com/news/getty-images-v-stability-ai-unpacking-the-high-courts-judgment), timeline of the EU AI Act (https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/)
• Business transfer platforms, listings displayed on 5 October 2026: CessionPME (https://www.cessionpme.com/), BusinessesForSale.com (https://www.businessesforsale.com/search/businesses-for-sale), BizBuySell (https://www.bizbuysell.com/businesses-for-sale/), Place des Commerces (https://www.placedescommerces.com/), BusinessBroker.net (https://www.businessbroker.net/about.aspx), RightBiz (https://www.rightbiz.co.uk/), DealStream (https://www.dealstream.com/businesses-for-sale), IndiaBizForSale (https://www.indiabizforsale.com/business/business-opportunities-for-sale), SEEK Business (https://www.seekbusiness.com.au/businesses-for-sale), CommercialRealEstate.com.au (https://www.commercialrealestate.com.au/business-for-sale/), Kumo (https://www.withkumo.com/), Bpifrance's Bourse de la transmission (https://reprise-entreprise.bpifrance.fr/), Bsale (https://www.bsale.com.au/businesses-for-sale), Daltons Business (https://www.daltonsbusiness.com/search/), BATONZ (https://batonz.jp/), nexxt-change (https://www.nexxt-change.org/SiteGlobals/Forms/Verkaufsangebot_Suche/Verkaufsangebotssuche_Formular.html), Fusacq (https://www.fusacq.com/), Brookz (https://www.brookz.nl/bedrijven-te-koop)
• Swiss business transfer platforms, listings displayed on 5 October 2026: LEEZ, Remicom (https://www.remicom.com/fr/dernieres-offres), Firm4sale (https://www.firm4sale.ch), Capital First (https://capitalfirst.ch/fonds-de-commerce-entreprises-vendre/), Nachfolgeportal (https://nachfolgeportal.ch/marktplatz), Firmo (https://firmo.ch/kategorie/firmenangebote/), companymarket.ch (https://www.companymarket.ch/), Relève PME presentation, May 2026 (https://www.fmb-ge.ch/wp-content/uploads/2026/06/Pres_RPME-FMB-27052026.pdf), Business Broker AG (https://www.businessbroker.ch/de/firma-kaufen/angebote), Nachfolger-Börse (https://nachfolger-boerse.ch/angebote)
• Company media outlets: John Deere's The Furrow (https://www.thedrum.com/news/2018/05/24/john-deere-the-og-content-marketer-how-its-123-year-old-magazine-endures), history of the Michelin Guide (https://www.finedininglovers.fr/article/histoire-guide-michelin), readership of Coopzeitung and Migros-Magazin (https://www.kleinreport.ch/news/wemf-zahlen-coopzeitung-und-migros-magazin-bleiben-unangefochtene-spitzenreiter-104919/), Red Bull Media House (https://www.redbullmediahouse.com/en/about-us/), HubSpot's acquisition of The Hustle (https://blog.hubspot.com/marketing/why-hubspot-is-acquiring-the-hustle), closure of Future (https://mediagazer.com/221201/p17)
• Business transfers in Switzerland, benchmarks taken from LEEZ's 2026 reference guide to acquisition financing and checked against their sources: CHDU on SECO's SME Portal (https://www.kmu.admin.ch/de/sorgfaeltig-geplante-nachfolgen-staerken-langfristig-die-schweizer-wirtschaft), KMU Next in Blick (https://www.blick.ch/wirtschaft/firmenuebergabe-experte-joerg-sennrich-sagt-was-es-zu-beachten-gilt-rund-ein-drittel-aller-nachfolgen-scheitern-id18208596.html), Zürcher Kantonalbank (https://www.zkb.ch/de/unternehmen/blog/firmen/nachfolge/firmenuebergabe.html), Centre Patronal, interview published by the FAE (https://fae-ge.ch/actu/releve-pme-catalyser-les-transmissions-pour-preserver-le-tissu-economique-romand/), Auditrium (https://www.auditrium.ch/blog/articles/aktiengesellschaft-verkaufen-in-der-schweiz), VZ, MBO case (https://www.vermoegenszentrum.ch/ratgeber/fachartikel/management-buy-out), commentary on judgment 4A_220/2013 (https://www.walderwyss.com/publications/1454.pdf), SG Nachfolge-Praxis, KMU Spiegel 2015 (https://www.sgnafo-praxis.ch/blog-25-kmufinanzierung/), Observatoire BCV des entreprises 2025 (https://mobile.bcv.ch/pointsforts/dans-le-canton/2026/observatoire-des-entreprises-2025/transmettre--fusionner--racheter.html)
• Product-led growth: Pendo glossary (https://www.pendo.io/glossary/product-led-growth/)
• Acquisition financing: LEEZ's 2026 reference guide, “Financer le rachat d'une entreprise en Suisse”; Basler Kantonalbank study on SME succession in north-western Switzerland (https://www.bkb.ch/de/die-basler-kantonalbank/medien/medienmitteilungen/2025/bkb-studie-nachfolgeplanung-von-kmu-in-der-nordwestschweiz-viele-unternehmen-brauchen-eine-nachfolge)

Guide written by Johan Bavaud, 2026 edition.

↑ Contents