GEO vs SEO in 2026: 0 to 66% AI Overviews in 3 Days, and What I Would Do in Your Shoes

SEO and GEO, the short definition (and the acronym trap)

SEO, Search Engine Optimization, is about getting a page to show up in a search engine’s results so a human clicks on it. The unit of success is the click. The signals have been known for twenty years: semantic relevance, technical performance, link popularity. The payback is measured in weeks, inside Search Console, and it is counted in sessions.

GEO, Generative Engine Optimization, is about getting a source cited by a generative engine: Google’s AI Overview, ChatGPT, Perplexity, Claude, Mistral. The unit of success is not the click, it is the citation. The model needs something else: extractable passages, attributed figures, named sources, structure. The payback is measured in days, and it is counted in mentions.

Now the trap. GEO stands for Generative Engine Optimization, not geolocated search. That is the mistake one of the French top 5 articles on this query makes: it spends half its text on business listings and recommends a tool Google shut down in September 2023. I am not naming it, there would be no point, but keep the filter in mind: if the word “local” shows up in the first three paragraphs, close the tab. For the rest of the vocabulary, we keep a SEO and GEO glossary up to date.

What happened in France on July 22, 2026

Google announced the rollout of AI Overviews and AI Mode in France on its French blog on July 22, 2026. On an informational query, a block of text written by a model sits above the classic results, with a handful of source links off to the side. The first blue link, the one we have been fighting over for fifteen years, does not disappear. It moves down.

That calendar detail explains why almost all the French-language content on “geo vs seo” is useless to you today: the top 10 articles were published or updated before that date. They describe a French market where the AI Overview was an American curiosity we watched from a distance. That market existed until July 22, 2026.

My measurement: 6,175 keywords, before and after

Here is the heart of the article, and the only part nobody can copy. I took fourteen Semrush Organic Positions exports on the google.fr database, covering fourteen domains from our own universe (SEO, marketing, WordPress), which is 6,175 unique keywords tracked between May 22 and July 28, 2026. For each keyword, Semrush records the SERP features present at the time of the reading. All I had to do was count. Before the numbers, the limits, because a figure with no scope attached is exactly what I hold against the nine articles next door.

A note on the three-day gap between the July 22 announcement and the first detection on July 25: it does not mean nothing was showing on the 22nd, it means the tracking database took three days to reflect the change. If you follow your rankings with a third-party tool, you have the same blind spot. On switch days, look at the SERP by hand.

By search intent: the AI targets knowing, not buying

Semrush assigns one or more intents to each keyword, so the shares below do not add up to 100%. The gap is no less obvious for it.

Another read of the same dataset, and an even more telling one: 86.53% of the 349 SERPs with an AI Overview carry an informational intent. (Hack The SEO readings, google.fr, July 25 to 28, 2026) If your traffic lives on explanatory content, guides and definitions, you are the primary target. If you sell, you have a stay of execution.

By difficulty: the easy long tail is not spared

This is the result that surprised me most. Received wisdom says the AI concentrates on big generic queries and leaves the low-competition long tail alone, the one everybody tells you to start with. On my readings, that is false.

The Featured Snippet, in the same window

While I was counting AI Overviews, I counted Featured Snippets. Same database, exactly the same limits.

The cross-tab is the sharpest result in the whole measurement. Across the 527 SERPs tracked after launch, 349 show an AI Overview with no Featured Snippet, 12 show a Featured Snippet with no AI Overview, and 166 show neither. On these readings, we have never seen the two together. Not once.

I weigh my words on the interpretation, because this is where everybody skids. The two cohorts show the Featured Snippet’s presence collapsing at the same time as the AI Overview’s presence explodes, with perfect mutual exclusion in the lower cohort. What I do not have is the paired tracking that would let me claim one took the other’s slot on a given keyword. The correlation is spectacular, the causation is not established, and an article selling you the second on the strength of the first is selling you a hunch dressed up as data.

The French numbers nobody cites

Here we leave my own measurement and move to published studies. All French, all dated, all with a named organisation and a sample size. This is the material the rest of the top 10 is missing, and it is what gives you the real framing: yes, the display shift has begun; no, the French market has not flipped.

That 1 in 134 ratio is the number I put on the table when a client explains that he wants to “bet everything on AI” in 2026. It is not an argument for doing nothing, it is an argument about sequencing. We broke down the rest of the market numbers, including the American and worldwide data, in our SEO and GEO statistics, which we update as we publish.

Four studies, four answers, and that is normal

44% at Médiamétrie. 48% at CREDOC. 57.1% at ARCOM. 59% at OpinionWay. Four serious organisations, four incompatible figures answering the same question: how many French people use generative AI.

Two explanations, and neither of them is “one of them is lying”. Declarative surveys overestimate compared to passive measurement: you remember trying ChatGPT, you declare it, the panel never sees it. And the definition of “using AI” shifts from one study to the next: once in the month, once in the week, on a service identified as such. The number that actually matters is none of the four, it is ARCOM’s: under 1% of time spent online. You can absolutely have tried it and changed none of your habits.

Three numbers everyone copies that are wrong

Probably the most useful section of this article if you have a budget call to make this week. The three most cited GEO statistics do not say what people make them say.

1. The “70% trust AI answers” credited to Gartner does not exist

Gartner published the opposite. On September 3, 2025, the firm announced that 53% of consumers distrust AI-generated search results, that 41% find AI overviews more frustrating than useful, and that 61% would like to be able to turn them off, on a sample of 377 people. That is the exact reverse of what you read everywhere.

The likely origin of the mix-up is a YouGov survey from July 2026, where the trust being measured was in search engines and not in AI: roughly 72% trust in search engines, against roughly 37% for AI assistants. The figure got flipped and reattributed along the way. Method transparency, since that is the subject of this section: the full YouGov report sits behind a signup wall, I only cross-checked those two values through consistent secondary citations, so I give them as orders of magnitude. What does not move is that the “70%” is circulating in the article ranked first for this query in France, credited to Gartner, with no link to anything.

2. The “76% of AI Overview citations come from the top 10” was revised by its own author

Ahrefs, who published that figure, brought it down to 37.9% on March 2, 2026, on a base of 863,000 SERPs and 4 million URLs, pointing to the arrival of Gemini 3 and the spread of query fan-out. BrightEdge finds separately that only around 17% of sources cited in AI Overviews also appear in the organic top 10. The 76% is still the most recycled figure on the web on this subject, and it is the argument for everyone explaining that ranking well is all you need to get cited. That has not been true since March.

3. The “Princeton +40%” does not mean what people make it mean

The +40.9% is real, but it holds for one method (adding citations), on one metric (Position-Adjusted Word Count), on a simulated generative engine. Exact reference: arXiv 2311.09735, published at KDD 2024, GEO-BENCH benchmark, 10,000 queries. On the subjective impression metric, the best lever tops out at +28%. And the real test on Perplexity, across 200 samples, gives +22%. It is still an excellent result. It is simply not “+40% guaranteed AI visibility”, which is what you read on just about every GEO agency sales page.

A bonus, because it comes up in every board meeting: the Gartner forecast of a 25% drop in search volume by 2026 dates from February 19, 2024, comes with no published methodology and no published sample, and has never been revised. Google wrote on May 19, 2026 that its queries had reached an all-time high. A 2024 forecast against a 2026 statement: I am not telling you which to believe, I am telling you that you cannot cite the first while ignoring the second. We keep the running list of corrected figures in our SEO and GEO statistics.

The 9 GEO levers, ranked by actually measured gain

Here is the table nobody publishes in full, even though it has been sitting in the Princeton study since 2024. Nine methods tested, nine gains measured, in order.

Take thirty seconds on that last row. The only negative lever in the study is the only one twenty years of SEO taught us to do. Keyword density, occurrences in the H2s, semantic variants repeated: everything a generation of tools flagged green produces the opposite effect here. The eight positive levers describe something very simple and very old: writing a good article, sourced, quantified, readable. GEO rewards journalism.

What the French-language corpus completely ignores

Four subjects appear nowhere in the nine top 10 articles, even though they are the four questions clients actually ask us.

The robots.txt for AI crawlers. GPTBot, ClaudeBot, PerplexityBot and Google-Extended are blocked line by line. What does not get said enough: blocking Google-Extended does not keep you out of the AI Overview, which draws on the classic search index, and blocking GPTBot removes you from ChatGPT’s answers without protecting you from much, since your content keeps circulating through other sites’ citations. It is a licensing call, not a security one.

The llms.txt file. No major generative engine has publicly committed to reading it as of today. It costs ten minutes, it breaks nothing, and it forces you to write a clean description of your site. We put a llms.txt generator out there for free for that. Drop it in, tick the box, move on. Anyone selling you a four-figure llms.txt engagement is trading on the confusion.

Answer volatility. Ask a generative engine the same question three times three days apart: you will not get the same cited sources. It is structural, and the consequence for reporting is direct, a single reading means nothing, you need a series. That is the logic behind our GEO score, which measures a page’s extractability signals rather than its presence on a given day.

Engine-by-engine differences. Google’s AI Overview leans on its index. Perplexity runs its own live search. ChatGPT combines training memory and browsing. Optimising “for AI” as though it were a single entity is optimising “for search engines” in 2005 while ignoring that Google and Yahoo did not run the same algorithm.

The big table: SEO vs GEO

The comparison you came for, with two columns I have not seen anywhere else: what you actually measure, and the French figure that goes with it.

My verdict

The two truths in this article contradict each other, and both are sourced. France is the least AI-dependent large Western market I have seen measured: 0.26% AI traffic, under 1% of time spent online, 98% of French people still using a search engine, less than 5% of retail traffic coming from AI agents. And at the same time, the display shift did happen, dated, in three days, in front of me.

Both hold because they are not talking about the same thing. One is about where the traffic comes from. The other is about what shows up above your results. And the second always comes before the first.

Frequently asked questions

Sources