There is a question that comes up almost every time in the first conversation with a company director, and it usually arrives towards the end, once the problem has been put on the table. It concerns AI Overviews, the summaries Google now displays above its results.
Is it still worth investing in SEO now that Google answers all by itself up there? The question is perfectly legitimate, and yet the answer is not really found where everyone usually looks for it.
AI Overviews: what the studies actually measure
Let us start with the serious figures, because a lot of loose ones have been circulating on LinkedIn for a year now. In July 2025 the Pew Research Center published a study based on the real behaviour of 900 people followed for a month, covering 68,879 analysed Google searches. When a generated summary appears, the click-through rate on a classic result falls to 8%, against 15% when it is absent. Roughly speaking, half the traffic evaporates.
And the links cited inside the summary itself? They collect 1% of visits. In other words, being cited by AI brings almost no one to your site, and that changes a great deal in the way some providers are currently selling the topic.
Seer Interactive went further in September, on a panel of 3,119 informational queries and 25 million impressions: a 61% fall in the organic click-through rate, and 68% on paid. The gap with Pew is explained simply enough, since Seer measures only pure informational search.
Google has in fact disputed Pew’s methodology, which is fair enough. Even so, six or seven independent studies have converged since 2024, with ranges varying from 15 to 79% depending on the method used. The direction, however, never varies.
Not every query triggers a summary
And now comes the point almost nobody raises, and yet it is the most useful one.
Not every query triggers an AI Overview, far from it. Pew measured this too: 8% of one- or two-word searches generate one, compared with 53% for those of ten words or more, and “why” or “how” questions climb to 60%. Short, transactional or local queries, on the other hand, very rarely trigger one.
In other words, the damage is concentrated on one very specific type of traffic: the one that was already converting the least well.
Take the case of a Lausanne fiduciary that we supported last year. Its prospects were obviously not looking for the definition of a balance sheet, they were looking for someone to entrust theirs to. As a result, the traffic it lost on definitional queries never represented its real commercial potential, it was volume, and volume is not the same thing as demand.
The first thing to do, and it costs nothing
Hence the first distinction to make, and it is simpler than it looks.
In practical terms, open Search Console, export your queries over twelve months, then split them into two piles: on one side those starting with a question word, on the other those containing a place, a price, a brand name or a comparison. Then look at where your completed forms actually come from, because that is where the answer lies.
With most of the SMEs we work with in French-speaking Switzerland, the result surprises the owner, since informational traffic makes up the volume and fills the charts nicely, without filling the order book. And if you do not know where to start with that export, our audit method using six free tools describes the complete sequence, tool by tool.
The real shift: search that starts in ChatGPT
That said, there is a real subject behind all this, and it is not where people usually expect it.
The Lucerne University of Applied Sciences measured in 2025 that 20% of the Swiss population had used ChatGPT or Copilot to look for a supplier in the previous twelve months, and, more importantly, that 13% started there directly, without going through Google. One person in eight, in Switzerland, not in Silicon Valley. That is the movement that genuinely matters, far more than the falling click-through rate on definitional queries. The governance of these tools is, incidentally, partly being written fifteen minutes away by tram, which does not hurt.
Except that, and I would rather tell you frankly than sell a certainty that does not exist: nobody yet knows how to optimise this reliably, nobody. There are converging observations, not an established method. What we see on sites that do get cited comes down to three points, which I give with their caveats.
Factual clarity counts for a lot: a passage that answers a precise question in under fifty words, let us say, gets picked up markedly more often than a nuanced three-hundred-word development. That is annoying for anyone who loves writing, but it is what we observe.
Then comes local anchoring, which counts for more than you might imagine. The models swallow mainly French-language corpus in which France weighs very heavily by volume, and so they naturally lean towards French sources. You therefore have to point their nose at it, by naming the cantons, by citing the LPD rather than the GDPR when that is the subject, or by writing amounts in francs. Without that precision, your content dissolves into a mass far larger than your own.
Attribution matters too, and this is the point most organisations overlook. An article signed by someone identifiable, with a role and a background that can be checked, gets picked up more readily than an anonymous text. This shows up fairly clearly in the data, even if nobody can say by exactly how much. Pew notes, however, that citations remain dominated by Wikipedia, YouTube and Reddit, which puts into perspective the hope of appearing there when you are a twenty-person SME.
What this narrative hides
That is precisely why I am wary of the prevailing discourse on the subject.
You hear all over the place that search optimisation is dead and that everything has to be rebuilt for artificial intelligence. The argument is convenient, because it justifies a billable redesign. But when only 1% of users click the links in a summary, optimising to appear there is a bet, not a strategy. A fast site, properly structured, with pages that answer a clear buying intent, that is still what makes the difference. And on that, AI Overviews have changed absolutely nothing.
What has changed, on the other hand, is how traffic is distributed. Informational search is eroding, transactional search is holding up, and the effort must therefore shift accordingly. This is exactly the kind of trade-off that a properly built SEO strategy has to set out in black and white, with figures and a schedule.
Where to start this week
In practical terms, and in this order.
First look at your PageSpeed score on mobile, page by page and not only on the homepage, because many French-speaking Swiss sites are weighed down by four-megabyte photos meant to look premium. Then pull from Search Console the split between informational and transactional, and identify the five pages that genuinely carry your acquisition, bearing in mind that they are rarely the ones you would expect. Work on those first.
In practice, allow two to three weeks for an SME, and with entirely free tools at that. What is most often missing is the discipline of looking at your own figures rather than listening to the last supplier who called, and that discipline cannot be bought.
Finally, a word about the figures quoted above, because it would be dishonest to present them otherwise. They come mainly from American panels, and search behaviour in French-speaking Switzerland is not identical, particularly on local and multilingual queries. They should therefore be taken for what they are, namely a solid trend, and not a measurement of your market.
