Google's advice to French brands is sound, and it stops where measurement begins. In an article signed by François Loviton, who runs Google's advertising solutions in France, the company sets out how to become the answer: say what only your brand can say, write for customers rather than bots, keep the website flawless, feed rich product data, and let AI-powered campaigns cover the conversational long tail. We agree with most of it. This article reads the same roadmap with the questions a brand has to answer afterwards: what the figures actually measure, what the answer means, and how a French company can know where it stands before its first peak season with AI Overviews.
Most figures in this debate come from Google itself, as internal data or studies it commissioned. We cite each one with its source and what it measures. KAIS adds no figures of its own to this article.
What changed in France on 22 July 2026?
Google switched on AI Overviews and AI Mode in France on 22 July 2026, more than two years after the global launch began in May 2024. Sébastien Missoffe, managing director of Google France, announced both at once. The wait was largely regulatory. Next reported that the launch had been held back by Google's obligations on neighbouring rights for press publishers, after the Autorité de la concurrence fined the company €250 million in 2024.
Three details of the launch matter to brands. AI Mode relies on query fan-out: it splits a question into sub-topics and runs many searches at once, which is why visibility is better measured by query territory than by keyword. Google says an AI Mode query is on average three times longer than a traditional one. And site owners in France got a new Search Console control that keeps a site out of generative AI features, which Google says is not used as a ranking signal for the rest of Search.
What does Google advise brands to do?
Google's roadmap rests on one premise: AI experiences use the same fundamental ranking principles as Search, so SEO best practice stays decisive. From there it draws three principles and a checklist.
The principles are to put forward what only your brand can express, with expert opinion, images, video and accurate product and local listings; to stop writing for bots, because the systems reward content that is genuinely useful; and to keep the website experience flawless, since what makes a site readable by an agent also makes it better for a person. The checklist adds reliable measurement (Google says advertisers adopting Google Tag Gateway observe 14% more conversions on average), product feeds enriched with conversational attributes for the Shopping Graph, local presence through Maps, which Google puts at 42 million monthly users in France according to Médiamétrie, and budgets flexible enough to absorb demand peaks. On paid search, the message is that the best ads now read like answers, with AI Max, Performance Max and Demand Gen reaching queries that keyword lists cannot cover.
Most of this is right, and some of it predates AI entirely. What follows is not a rebuttal. It is the rest of the question.
What do the figures in Google's argument actually measure?
Each figure measures something narrower than the sentence around it suggests. None is false. But an exposure count is not usage, a declared feeling is not a behaviour, and a US retail conversion rate is not a French AI Overviews result.
| Figure | Source | What it measures | What it does not say |
|---|---|---|---|
| 2.5 billion monthly users of AI Overviews | Google, I/O keynote, May 2026 | People who saw an AI Overview in a month | Exposure, not chosen use |
| More than 1 billion monthly users of AI Mode | Google, I/O keynote, May 2026 | Use of a dedicated interface people open on purpose | Global figure declared by Google, not audited |
| AI Mode queries three times longer | Google, US data | Average query length | US behaviour, not yet measured in France |
| 16 searches per journey with AI, against 8 without | Google, Mastering the Messy Middle, cited in its article | Searches per purchase journey across 7 European markets | Google study whose full method we could not find published |
| 3 users in 4 decide faster and with more confidence | Ipsos for Google, December 2025, 25-market average | What users say about their decisions | Declared, in a study commissioned by Google |
| 42% higher conversion | Adobe Analytics, US retail sites, March 2026 | Visits arriving from AI sources against non-AI channels | Does not isolate AI Overviews; 38% lower a year earlier |
| 8% of visits with a click when an AI summary appears, against 15% without | Pew Research Center, US, March 2025 | Clicks on traditional results | An association, not a proven cause |
Two of these deserve a closer look. The 42% comes from Adobe, which tracks more than a trillion visits to US retail sites: in March 2026, visits arriving from AI sources converted 42% better than non-AI traffic such as paid search and email. In March 2025, the same traffic converted 38% worse. The figure describes people who clicked through from AI assistants and chat services. It does not isolate users of Google's AI Overviews, and it cannot tell whether AI makes shoppers more decided or whether decided shoppers are the ones using AI.
The 2.5 billion counts everyone who saw an AI Overview in a month, whether or not they read it. AI Mode's billion is closer to chosen use. Pew Research Center's browsing study adds the part Google's figures leave out: in March 2025, US users who saw an AI summary clicked a traditional result in 8% of visits, against 15% without one, and clicked a link inside the summary in 1%. Pew reports an association, not a cause. It is still the clearest public sign that being visible inside the answer and receiving traffic from it are now two separate outcomes.
Is ranking well in Google enough to appear in AI Overviews?
Ranking makes a page eligible; it does not make it the answer. Google is right that the same fundamentals apply: a page that is not indexed or not trusted will not be used. But an AI Overview keeps a handful of sources among the pages it could use, and AI Mode goes further, fanning a question out into sub-queries that a single keyword ranking never tracked. We have documented how first place on Google and citation in AI answers have drifted apart.
The answer is not one thing either. A brand can be cited as a source, named as a provider or recommended as the solution, and those three outcomes can move in opposite directions. Google's roadmap treats appearing as the goal. For a company that sells, the goal is the recommendation, and it is measured differently, as set out in our analysis of Recommendation Share of Voice.
Why does AI make people search more, and what does it change for brands?
Because an AI suggestion invites a check, and that check is now answered by AI as well. Google's own research makes the first half of the point. Journeys that involve AI include more searches, not fewer: 16 on average against 8 without AI, according to its Mastering the Messy Middle work. And searches asking whether a brand is reliable are rising across Europe in Google Trends. Google's conclusion is that people buy from brands, not from algorithms, so brands must build the reassurance that completes the AI's sorting.
The second half is where that reassurance now happens. When a buyer asks whether a brand can be trusted, an AI Overview can answer directly, by summarising reviews, forums, press coverage and complaints. The reassurance a brand builds on its own site becomes one source among several, and often not the one quoted. What the engine says about a brand when asked to vouch for it has become part of the decision.
That gives two moments to measure, and they should never share a score. The selection moment covers generic questions that name no brand, where the engine builds the shortlist; the measure there is recommendation. The verification moment covers questions that name the brand, where the engine confirms or undermines the choice; the measure there is accuracy, tone and the sources the engine relies on. KAIS keeps branded questions out of visibility scores, because a query that names a brand always finds it. That is exactly why they need an audit of their own.
What is a product feed for a company that sells services?
For a service business, the feed is the entity: a precise, consistent description of what the company does, where, for whom and on what evidence. Google's checklist is written for retail, and its core idea travels well: an AI does not recommend what it does not understand. Google's example contrasts a bare product name with one that spells out colour, fabric and occasion. For a law firm, the same gap separates "corporate law" from "setting up a Thai company for a French founder, in French, with fixed fees".
In practice that means services named the way clients ask for them; addresses and opening hours that match everywhere; practitioners with verifiable credentials; price ranges where the profession allows them; a complete Business Profile, since Maps is where local intent lands; structured data stating the same facts; and third-party profiles that say the same thing. It also means a site that machines can actually read. Adobe's audit of US retail found product pages only 66% readable by language models on average, and a page can be public and still unreachable to AI crawlers because of a firewall.
Is an ad inside an AI answer worth a recommendation?
No. A sponsored placement is bought and labelled, a recommendation is earned, and a report should never add the two together. Google's article argues that the best ads are now answers, written in natural language to explain how a product solves a precise problem. That is good copywriting advice, and AI Max, Performance Max and Demand Gen do reach conversational queries that keyword lists miss.
But paid and organic presence answer different questions. The ad tells you what reach costs. The organic mention tells you whether the engine, looking at the evidence available about your market, puts you forward. A brand that only appears in the sponsored slot has bought visibility inside an answer that recommends someone else. Both can be worth having. They belong on separate lines of the report, with separate budgets.
Should a brand use the Search Console control that keeps a site out of AI features?
Only as a deliberate decision, because a brand that opts out removes its own site from the material the answers are built on. The control launched alongside AI Overviews in France lets a site owner decide whether the site appears in generative AI features and contributes to their answers. Google states it is not used as a ranking signal outside those features. It makes sense for a press publisher weighing neighbouring rights. For a company that wants to be recommended, switching it on means the engine will describe the company from other people's pages only. That decision belongs to management, in writing, and someone should check that the setting stays the way it was decided.
How can a French brand measure its visibility in AI Overviews?
By observing the answers directly, because Google's tools report impressions, not answers. The generative AI report in Search Console, rolled out to all properties on 31 August 2026, shows impressions of your own pages only, with no queries, and counts AI Overviews and AI Mode together. No Google tool shows competitors.
The protocol we use is observational. A frozen panel of questions, grouped by intent and territory, with branded questions audited separately. AI Overviews and AI Mode read as two surfaces, because they do not cite the same sources. Each question run three times on recorded dates before any gap is treated as real. Each answer coded by role: cited, named or recommended. The method is detailed in How to compare your visibility to competitors in AI Mode.
Why does peak season 2026 call for a baseline now?
Because it is the first French peak season with AI Overviews and AI Mode, and without a reading taken before it, nobody will be able to say what they changed. US brands have had two holiday seasons with AI Overviews to compare. French brands have none. Google's closing advice, to stay curious, test and iterate, assumes a reference point to test against.
A baseline is simple, and it cannot be taken after the fact. Freeze the question panel in September or October, read AI Overviews and AI Mode on dated runs, record roles and sources, then read the same panel again during and after the peak. Without that first reading, any movement in sales, traffic or branded search will be credited to AI, or dismissed, on opinion. With it, the discussion rests on what the engines actually said.
Where does each of Google's recommendations sit in the KAIS AI Visibility Chain?
Google's roadmap covers the early stages of the chain well and leaves the last two unmeasured.
Be Discovered → Be Retrieved → Be Understood → Be Trusted → Be Cited → Influence the Answer → Be Recommended.
| Google's recommendation | Stage of the chain | What to measure |
|---|---|---|
| A flawless site, readable by agents | Be Discovered, Be Retrieved | The response code and content AI crawlers actually receive |
| Unique content, expert opinion, accurate listings and feeds | Be Understood | Whether the engine describes the offer correctly |
| Reassurance for verification searches | Be Trusted | Accuracy, tone and sources on questions naming the brand |
| SEO best practice | Eligibility to Be Cited | Share of source citations on the question panel |
| Not addressed in the roadmap | Influence the Answer, Be Recommended | Recommendation rate and share of recommendations against competitors |
The mapping explains how sound fundamentals can coexist with weak results. A site can be reachable, understood and trusted, cited often, and still rarely recommended when a buyer asks whom to choose. The chain also has a documented exception: an engine can recommend a brand from what its model already learned, with no visible source. That is one more reason to read the answers rather than count citations.
Being the answer, and knowing it
Google's article ends on a posture: stay curious, test, iterate, and bet on what consumers look for, which is knowledge, authenticity and trust. We would add one discipline. Know which answer you are in, in what role, next to whom, and how that has changed since the last reading.
French brands start this peak season with no history. That is a handicap for this year's comparisons, and an advantage for every season after it, for the brands that take their first reading now.
Sources
- Google France — Lancement des Aperçus IA et du Mode IA dans la Recherche Google en France, Sébastien Missoffe, 22 July 2026. Launch date, global rollout since May 2024, query fan-out in AI Mode, queries three times longer, the Search Console control for AI features.
- Google (Think with Google France), article by François Loviton on AI Overviews and peak season, 2026. Google's roadmap for brands and the figures it attributes to its internal data, to its Mastering the Messy Middle report, to Ipsos (study commissioned by Google, December 2025) and to Médiamétrie for Maps.
- Google — Google I/O 2026: Sundar Pichai's opening keynote, May 2026. More than 2.5 billion monthly users of AI Overviews and more than 1 billion of AI Mode.
- Adobe — Adobe report: U.S. retailers see surge in AI traffic, but many websites are not entirely readable by machines, Vivek Pandya, 16 April 2026. Conversion of AI traffic (42% better in March 2026, 38% worse in March 2025) and machine readability of retail pages.
- Pew Research Center — Google users are less likely to click on links when an AI summary appears in the results, 22 July 2025. Browsing data of 900 US adults, March 2025.
- Next — AI Overviews : Google lance ses Aperçus IA dans la recherche en France, 22 July 2026. Context of the French delay and neighbouring rights.
- Google (Think with Google UK) — Your 2026 marketing strategy starts with peak season data, 20 January 2026. Average conversion uplift of 14% observed by advertisers adopting Google Tag Gateway.