An executive showed us his screen during the first conversation, at the start of the year. He had asked ChatGPT which engineering firm to recommend for a building renovation in Fribourg. The answer cited two competitors and a specialist directory, but not him. His question was one sentence long: why them? Yet most of the answer lies in the types of content those sites publish, and rather little in their size or reputation.
That is worth pausing on, because it means an SME can play this game too. Provided it understands what these machines actually read, and what they leave aside.
What the machine looks for in your content types
A language model does not read a page the way a visitor does. It cuts it into passages, keeps what is stated clearly, then assembles an answer from the most usable passages. As a result, a paragraph that states a dated, quantified and attributable fact is easy to reuse. A long, cautious development with no figures and no stance gives it nothing to extract, however well written it may be.
That is why certain formats keep coming back in generated answers. An explicit definition at the top of a section. A table when there is a comparison. A short answer placed just below the question it addresses. It is not a mysterious taste for lists on the algorithm’s part, it is extraction: what cuts cleanly, cites cleanly.
The technical layer counts too, and we set it out in detail in how to structure a site so that AI cites it correctly. Simply remember that clean Schema.org markup tells the machine what it is looking at, a price, a review, a step, without it having to guess.
The study that put a figure on the phenomenon
Until 2024, all of this came down to field intuition, it has to be said. A team from Princeton, Georgia Tech and IIT Delhi then published the first serious measurement, presented at the KDD 2024 conference under the name GEO, for Generative Engine Optimization. The researchers ran some 10,000 queries and compared variants of the same page against answer engines.
Three changes stand out: adding sourced statistics, adding expert quotes, adding verifiable references. On their visibility metric, these additions improved presence in answers by 30 to 40%. Improving the flow of the text added a further 15 to 30%. Conversely, keyword stuffing, the old SEO recipe from fifteen years ago, performed about 10% worse than the original page on Perplexity.
I give these figures with their caveat, because it is a real one: the experiment pits five sources against each other in a controlled environment, and the gains are relative. In a competitive market, the effect will be more modest. The most useful result is actually further down the study: pages ranked fifth on Google, almost invisible in generative answers, gain 115% visibility once enriched with references. In other words, the machine does not respect the hierarchy of positions as much as we imagine. That is precisely where an SME has a card to play, since it will never occupy the top position on the big queries.
Who actually gets cited
Let us stay clear-eyed about the starting point, all the same. In July 2025 the Pew Research Center analysed the actual behaviour of 900 people over a month: the sources most cited in Google’s summaries remain Wikipedia, YouTube and Reddit. A twenty-person company will not unseat Wikipedia on “what is a second pillar”. That battle is lost in advance, and that is fine.
On trade-specific and local questions, on the other hand, there is almost no one. Try looking for who explains how much LPD compliance costs for a ten-person accounting firm, or where to start when renovating the envelope of a building in Fribourg. Three sites, sometimes two, often none. There, a well-built page becomes the available source, almost by default you could say. How much traffic that brings in afterwards is another story: we have looked at the question with the figures to hand, and the answer calls for caution.
The content that makes the difference
First comes proprietary data. Your price ranges as observed, your real lead times, an anonymised but properly situated client case, canton included. Researchers call this information gain: a page that repeats what fifty others already publish gives no reason to be cited. Your own figures, nobody else has them, and that is precisely why crafted content beats generated content.
Then comes the page that answers a single question, the one a client actually asked you. The questions you receive by email are worth more than any keyword tool, since they arrive already phrased the way people type them into ChatGPT. A clear answer under the heading, the development after, the nuances at the end of the page.
The classic example among our clients is the “pricing” page with no price on it. No machine will cite it, there is nothing in it. Give a range, even a wide one, even with all the caveats you like. It will always be better than a silent form.
Then there is the byline, the real one. A text attributed to an identifiable person, with a verifiable role and background, gets picked up better than content signed “Admin”, and by a wide margin. The same goes for Swiss anchoring. The French these models have swallowed comes overwhelmingly from France, so when nothing signals otherwise they answer as if you were French: auto-entrepreneur, GDPR, prices in euros. It is up to you to put the right markers in front of them, in black and white. The LPD when it is the LPD, amounts in francs, canton names. Otherwise your page files itself away in the wrong drawer.
What I am not promising you
In 2025, an independent benchmark called C-SEO Bench went through, one by one, the “conversational” tactics circulating on LinkedIn. The result: most change nothing measurable. A few even perform worse. What holds up, again and again, is that the page genuinely answers the question asked. One would like a better-kept secret, but there is none.
Nobody can guarantee a citation in ChatGPT, nobody. Anyone promising you otherwise is selling a bet, not a service. What can honestly be built is pages that extract cleanly, data that only you hold, and a verifiable signature behind it. The rest belongs to the models, and they change without warning.
Where to start this week
Open ChatGPT and Perplexity, then put to them the five questions your clients ask you most often, in their own words. Note who gets cited, every time. The exercise takes twenty minutes and it usefully replaces a good deal of internal debate about “our AI visibility”. Keep a dated record of these answers, in a simple table. The models change nearly every quarter, and it is the comparison over time that will tell you whether your work is paying off. Not an isolated screenshot.
Then take your most visited page and give it what is missing: a dated and sourced figure, a short answer under the title, a table if the page compares anything at all. None of this requires a redesign, or even a budget as such.
Finally, if the exercise reveals that your competitors are already occupying the ground, this is the moment to talk about it. A serious SEO strategy now includes this generative dimension, with figures and a timeline, and the first conversation is a chance to see whether the subject deserves a project at your end.
