AI SEO: what the term actually covers in 2026

Ask five people what "AI SEO" means. You will get five answers: getting cited by ChatGPT, writing articles with a model, showing up in AI Overviews, plugging agents into Search Console, or simply "doing SEO in 2026". The term has become a commercial catch-all.

Rather than adding one more definition, we measured. 162 French keywords pulled from DataForSEO on 2026-08-27, and twelve months of Search Console for vydera.com re-exported with full pagination: 1,579 queries and 8,706 query plus country pairs.

The result drives everything else in this article, and it is not the one we expected: in French, people search for the name of the thing, never for the work. The vocabulary that names AI SEO has volume. The vocabulary describing what you do there has none. And that action vocabulary shows up intact in Search Console, typed no longer by humans but by engines.

Three crafts under one word

Before the numbers, the sorting. "AI SEO" covers at least three things that are not measured the same way and are not worked with the same levers.

• Being cited by generative engines. Called GEO or AEO depending on the school. The goal is to be the source a model retrieves and attributes. The click is not the measurement.
• Existing inside a Google that answers by itself. AI Overviews and AI Mode change the surface, not the index. The work stays SEO, but the outcome reads as impressions without clicks rather than as traffic.
• Producing SEO with AI. Agents, prompts, audit automation. That is an internal tooling topic, unrelated to how visible the brand is inside answers.

The confusion is not merely semantic, it gets invoiced. An executive commissioning "AI SEO" with the first sense in mind can be handed the third: articles produced faster, on the same queries as before. Both services are legitimate, they simply have nothing to do with each other. So the first question to put to a provider is not "do you do AI SEO", it is "which of the three, and measured with what instrument".

The three senses live in the same lexical field, and our measurement confirms it: of the 77 keywords we classified as GEO, 33 belong to visibility and citation, 18 to engines and platforms, 16 to AI-assisted production, 10 to crawler access. Four families, four crafts, one word to sell them all. For the boundary between the acronyms, we drew it in our GEO and AEO definition and in our SEO versus AEO comparison.

Seven surfaces, and what we can honestly measure on them

The useful question is not "what is AI SEO", it is "what plays out on which surface, and what do I actually see of it". Here is the map, with what we measure on each and what we do not.

That map is uncomfortable, which is exactly why we publish it. Five surfaces out of seven are instrumented nowhere on our side. No citation of vydera.com inside a ChatGPT, Perplexity, Gemini or Copilot answer has been observed, because nothing observes it. Everything we read about those surfaces is a retrieval signal in Search Console, never a verified citation. An agency that quotes you its citation rate without naming the instrument is selling you an estimate.

What the map changes in practice is the unit of success. On Google organic it is called position, and a rank gained converts into clicks. On AI Overviews and AI Mode it is called retrieval: the page served as a source, the visitor may never arrive. On the four generative engines it is called citation, and nobody sees it without dedicated instrumentation. Steering those three surfaces with a single indicator, the click, means writing off most of the work as useless.

The naming vocabulary has volume

On to the lexical field. Of the 77 keywords in the GEO field, 26 reach 20 searches a month or more. They are almost all nouns: the name of a file, the name of a discipline, the name of a bot.

Four observations on that table.

One. The biggest term in the field is "ai overview", 2,900 searches a month. That is a definition query, not an optimisation one: people want to know what it is, not how to get into it. We make no comment on its trend, since twelve months cannot separate a news spike from a rising curve.

Two. The other 51 keywords in the GEO field sit below 20 searches a month, meaning two thirds of the field. Our own strategy notes said "nearly all of it". That was wrong, and we correct it here: two thirds is not nearly all. The practical conclusion does not move, but a loose number has no place in a document used to make decisions.

Three. The "engines and platforms" sub-family is the black hole of the field. It holds 18 keywords, of which a single one, "ai overview", carries the entire volume. The other 17 add up to 60 searches a month, and 12 of them return no data at all. "être visible sur chatgpt", "optimiser pour perplexity", "comment être cité par chatgpt": nobody types that in French.

Four. The GEO field is not narrower than the classic SEO field, contrary to what we would have bet. Concentration is comparable: it takes 8 keywords to cover 80% of GEO volume against 9 on the classic SEO side, and the technical family is the most concentrated of the four, with 79% of its volume on its top 5 terms.

The operational reading of the table sits in two columns read together: volume and difficulty. Six terms in the field show a difficulty of 0 with usable volume: "claudebot" and "référencement ia" at 390 searches a month, "query fan out" and "seo vs geo" at 320, "agent ia seo" and "citation ia" at 140. Those are the realistic entry points for a young domain. At the other end, "generative engine optimization" carries 880 searches and a difficulty of 39: that is a six-month objective, not one more article.

One methodological point, because it changes the reading: a missing volume is not a zero. It is a term below the reporting floor of Keyword Planner. We never add it as a 0, and we count it as sitting below the threshold of 20, which is true by construction.

The vocabulary describing the work has none

Here is the result we did not expect, and the only honest way to test it.

The first split between noun-shaped and action-shaped keywords ran on hand-written terms. Immediate objection: an invented phrasing returns nothing because nobody typed it, not because action-shaped demand is invisible. The test proved nothing.

So we ran it the other way round. 22 queries pulled from the French Search Console of vydera.com, therefore actually served by Google to French users, fed back into Keyword Planner exactly as they were. Same origin, same instrument, one variable: the shape. 10 phrased as an action or a question, 12 phrased as a noun.

0 of the 10 action-shaped phrasings return a volume. 6 of the 12 noun-shaped ones do. And across the whole set, 16 of the 22 queries Google actually served are invisible to the planner.

The example to remember is "meilleurs outils geo": 16 impressions and 1 click in France over twelve months, average position 27.1, and no volume in Keyword Planner. The click proves a human typed it. The tool says it does not exist.

Two checks were needed, and we publish both. "netlinking geo" served as the example first, then got dropped: 212 impressions at an average position of 9.3 with 0 clicks does not look like a human query, it looks like fan-out that the word-count rule fails to catch. And the positions of all 12 noun-shaped queries were measured for the same reason: the zero CTRs in that group are explained by position, 42.0 for "outil geo", 46.4 for "seo claude", not by a synthetic origin.

One limit to keep in mind: the classic SEO field holds only 2 action-shaped keywords in our set of 162. That is far too few to serve as a control group. We cannot write that the shape effect holds whatever the topic. It is established on the GEO field and on those 22 observed queries, not beyond.

Where the demand went: it became fan-out

If nobody types "how do I get visible on ChatGPT", has the question disappeared? No. It changed keyboards.

In our Search Console, in France, over twelve months: 40 action-shaped or question-shaped queries, 244 impressions, 0 clicks. 36 of those 40 run to 6 words or more. The same vocabulary is empty on the planner side and clickless on the Search Console side. This is not demand dying out, it is demand that moved from the user to the engine, which manufactures it as sub-queries to go and fetch sources. That is the mechanism we detail in our article on query fan-out.

The 6-word threshold was not chosen, it was observed. Here are the three tiers, and the second one is the one we would rather have kept quiet.

• 1 word: 51 queries, 1,102 impressions, 353 clicks, CTR 32.0%. That is the brand name. On its own it accounts for 86% of the corpus clicks.
• 2 to 5 words: 682 queries, 28,352 impressions, 56 clicks, CTR 0.198%. The real non-brand traffic.
• 6 words and up: 846 queries, 13,962 impressions, 2 clicks, CTR 0.014%.

Comparing the overall CTR of short queries, 1.389%, against long ones at 0.014% would give a flattering and wrong factor of 99: it mixes the brand in with everything else. The honest ratio sits between the second and third tiers, 0.198% against 0.014%, a factor of 14. The length effect holds, it is simply fourteen times smaller than a shortcut would suggest.

The country comparison is sharper still. In France, 490 queries, 4,200 impressions, 296 clicks, CTR 7.05%. In the United States, 939 queries, 20,874 impressions, 11 clicks, CTR 0.05%. Two visibility regimes on the same site: humans clicking on one side, an engine retrieving on the other. Fan-out accounts for 27% of French impressions and 40% of American ones.

Spotting fan-out in your own Search Console takes three minutes. Export the queries, count the words, then look at the shape of those running past six words: if they read as sentences, with a conjugated verb and sometimes a question mark, they are not searches, they are manufactured sub-questions. One of ours runs to 26 words and opens with "j'ai testé un outil geo qui ne donne que des dashboards". Nobody types that into a search box.

On the French GEO field as seen by Search Console, the verdict is blunt: 386 queries, 2,506 impressions, 16 clicks. Including 207 synthetic queries for 1,022 impressions and a single click.

The caveat that stops this being a law. Our long-query corpus is 59% the fan-out of one single article: 222 of the 846 queries contain the name of a brand featured in a comparison piece, and they carry 8,198 of the 13,962 impressions. Removing them leaves 624 queries, 5,764 impressions and 2 clicks. The CTR collapse holds without them, but the fact describes the fan-out of vydera.com, not a general law about query length.

Second caveat, on the rule itself: the only French click our marker-based variant files as synthetic is "claude vs perplexity vs chatgpt", 4 words, 1 impression, 1 click. A human that the "vs" marker misfiled. That is why word count stays the primary measure and markers a variant published alongside, not a correction.

Volume says nothing about value

One last number, and it should give pause to anyone picking targets by volume. "visibilité ia" is worth 17.42 € per click at 50 searches a month. "ai overview", the most searched term in the field at 2,900, is worth 1.80 €.

Across the 20 keywords in the set that carry both a volume and a CPC, the rank correlation between the two is 0.12. That is not a weak relationship, it is the absence of one. You cannot conclude that the most searched terms are the cheapest, nor the reverse: the extremes suggest it, the whole set denies it, and n = 20 carries no test value.

What can be said fits in one sentence: picking a GEO target by volume means picking at random on the axis that matters.

What this changes in a strategy

Five operational consequences, in order.

1. Separate two engines and measure them apart. Capture human traffic on the naming vocabulary, and get retrieved on the action vocabulary. The first is judged in clicks, the second never is.
2. Target the nouns, write for the actions. Title and URL take the term that carries volume: llms.txt, generative engine optimization, AI SEO. The body answers the action-shaped questions, because that is the shape fan-out goes looking for.
3. Do not build an editorial plan on engines and platforms. Seventeen keywords, 60 searches a month. One framing article is enough, not thirty.
4. Never judge a citation asset by CTR. An article at 13,962 impressions and 2 clicks is not a failure, it is an asset working on the other engine. Provided that was decided beforehand, not afterwards.
5. Instrument what you claim to steer. We do not measure our citations, and we write it down. An agency that does not say so probably does not measure them either.

Point 2 is the most counter-intuitive, and the one that pays. A page optimised on the noun and written for the action wins on both counts: it ranks on the term that carries volume, and it holds the self-contained passages an engine will retrieve when it manufactures "quelle plateforme geo combine monitoring et création de contenu". Same page, not two.

Two technical prerequisites sit upstream of all this: letting in the bots you want citing you, covered in our article on AI crawlers and robots.txt, and serving an llms.txt file that actually stands up.

Running the measurement yourself

Nothing here needs an expensive tool. The two DataForSEO calls behind this measurement cost 0.18 $ in total.

1. Export your queries with full pagination, query and country dimensions, over twelve months. Watch the trap: our first export stopped at 1,000 rows, sorted by clicks then alphabetically. It was missing 579 queries and 8,941 impressions, including every accented French query, because "é" sorts after "z". Every sum drawn from it was wrong.
2. Count the words in each query. Look at where CTR collapses on your own corpus. The 6-word threshold is ours, not a constant.
3. Isolate the brand. Without that, your average CTR only measures how well known you are.
4. Take twenty actually served queries, half nouns, half actions, and run them back through the planner. It is the cheapest test available and the one that teaches the most.

One perimeter limit to know before publishing anything: the query dimension only covers 411 of the 1,371 clicks and 43,416 of the 109,907 impressions on our property. Google anonymises the rest. No query-level total can be presented as a site total, on our side or yours.

For the definitions: query fan-out and AI Overview.