AI SEO agency: 15 questions, and why 14 of 27 quote someone else’s numbers

Every agency website looks the same. Same vocabulary, same engines quoted, same promise of visibility inside ChatGPT and Perplexity. Screening therefore does not happen while reading a site, it happens in the meeting, question by question.

To find out which questions hurt, we measured the ground rather than imagining it. 32 French SEO and marketing agency domains probed on 28 August 2026, 27 measured, on their public HTML alone: homepage, service pages, dedicated AI pages, client cases, pricing, method. No blog and no guides: we only read the commercial shopfront, the one that commits the agency when a prospect reads it.

The starting hypothesis was the most banal in the industry: plenty of agencies promise AI and prove nothing. It is wrong. Of the 26 agencies that talk about AI, 8 show no verifiable evidence at all, or 30.8%. That is a minority, not a market built on air.

The real flaw sits elsewhere, and it cuts deeper. 7 agencies out of 27, or 25.9%, publish a number committing their own work inside a sentence about AI. 14 out of 27, or 51.9%, only quote market statistics borrowed from third parties. The French market is not silent about its results: it puts somebody else's number where its own should be.

What we measured, and how

The sampling rule is fixed and published, because it materially changes the result. For each domain: the homepage, then at most 8 service pages, 3 dedicated AI pages, 3 client case pages, 3 pricing pages and 3 method pages. Service pages are sorted first by the ones the agency links from its own homepage, then by increasing depth, then alphabetically.

That sorting is not decorative. Switching to alphabetical order alone filled one large site's sample with local long-tail pages and dropped its count from 309 to 130 hits. The rule lives in the script, and the whole survey recounts without reopening a single page thanks to the 117 MB of raw HTML kept on disk.

The URL pool comes from the homepage and the sitemap, up to 14 sitemaps opened per domain. Without the sitemap, a JavaScript-built menu leaves no link in the HTML, and three sites in the panel collapsed to their homepage alone.

Three definitions carry everything else:

• An own number is a figure committing whoever writes it: our clients, our timeline, from this amount onward. A figure attributed to a third party, "according to a study", is a market number.
• An execution tool is a piece of software or a script named in an AI context. A target engine is not one: quoting ChatGPT is a promise, not evidence.
• Evidence is one of three things: an own number, a named execution tool, or a quantified client case tied to AI.

The panel comes from the 24 candidates of our E-E-A-T audit, plus 8 domains probed with curl and kept on a 200. digidop.com is a company related to Vydera: it sits in the panel, it goes through exactly the same grid, and the tie is declared here so nobody mistakes this panel for a neutral sample. vydera.com is in there too, same rules and same failures.

The borrowed number in place of the owned one

96.3% of the measured agencies mention AI on their shopfront. Only one never does. So the topic is shared by the entire market, which teaches nobody anything.

What does teach something is how the numbers break down. Across 27 sites:

• 7 publish at least one figure committing their own work in an AI context: a timeline held, a client result, a starting price.
• 14 only quote market figures: a share of searches, a sector growth rate, a percentage lifted from a study signed by somebody else.
• 6 quote no figure at all in an AI context.

Substitution is the dominant behaviour, and it is comfortable: a market figure cannot be contradicted by an unhappy client. A promised timeline can.

It is also the fastest question to ask in a meeting. "What number do you publish about your own work?" If the answer opens with "according to a study", you have your answer in eight seconds.

The engine is named, the tool never is

Second substitution, same nature. 22 of 27 agencies name at least one target engine: ChatGPT, Perplexity, Gemini, Copilot, AI Overviews. That is the entry-level vocabulary and it costs nothing.

13 of 27 name at least one execution tool in an AI context. And 9 of the 22 that name an engine name no tool at all: they say where they want to make you visible without ever saying what with.

Across the whole panel, only thirteen distinct tools are named in an AI context: Semrush, Ahrefs, Screaming Frog, Semji, Similarweb, Python, n8n, Zapier, Looker Studio, Otterly, Scrunch AI, the Anthropic API and Claude Code. 11 of 27 sites name something other than Semrush or Ahrefs.

That nuance matters, and it pushed two sites into the no-evidence column: quoting Semrush on a classic SEO audit page is not evidence of AI execution. The tool has to be named inside the context of the promise, otherwise it is a logo in a carousel.

The 15-question grid

The fifteen questions below are the ones we ask in a meeting. Each maps to one axis measured across the 27 sites, and the counter on the right shows how many of them clear the same axis. A question whose answer we cannot measure is not in the list.

Two calibration results, taken from the panel itself. The median score is 9 out of 15, the mean 8.1. A median agency talks about AI, names engines, describes a method on its shopfront, and publishes neither its own number nor a price.

And above all: the highest score observed is 13 out of 15, reached by three sites. None of the 27 clears 14 or 15 axes. A grid nobody fills is not a broken grid, it is a grid that was not cut to flatter its author.

The panel, row by row

All 27 sites, sorted by score then by how loudly they talk. The columns carry the measurement, the conclusion is yours.

The most useful reading of this table is the disappointing one. Loudness does not predict the score. The three most AI-saturated shopfronts score 13, 11 and 7 out of 15 respectively. Conversely, a site at 1.96 hits per thousand words reaches 11. Talking a lot about AI and showing what you do with it are two independent skills.

The other transparency figures follow the same slope. 20 of 27 sites publish at least one quantified client case, any topic, but only 6 tie one to AI. 18 have a pricing page in their URL pool, 6 write an amount on it. 11 expose a methodology URL, yet 22 actually describe a method in the copy of their shopfront: the dedicated page is missing more often than the method itself.

What the grid says about vydera.com

Our own site is in the panel and goes through the same grid. It comes first on density, at 49.52 hits per thousand words, which is a fact about volume of talk and nothing more. It reaches 13 out of 15, tied with two other sites, and it fails on two axes.

First failure: a single execution tool named in an AI context, Claude Code, while other tools are named elsewhere on the site, outside the context where they would count. Second failure: no dedicated pricing URL. The price is quantified in the body copy, which clears question 13 and misses question 14.

There is a third point worth publishing, less visible. vydera.com quotes 27 market figures against 12 of its own. That is exactly the slope described above, at a less bad ratio than the panel average. A better ratio is not a different behaviour.

Two agencies that do not sell AI, and four brands that are only three

Two sites in the panel deserve to be cited for what they are, not for what they lack. korleon-biz.com mentions AI once across 11,239 words read. nextlevel.link never mentions it, across 2,149 words. These are not measurement failures, they are two agencies that do not sell AI. In a market where 96.3% of shopfronts talk about it, that is a position, and a perfectly defensible one.

Another lesson, structural this time. Of the 32 domains probed, 4 redirect to another brand already in the panel: mediaveille.com serves yumens.fr, 1ere-position.fr and open-linking.com both serve noiise.com, search-foresight.com serves peakace.fr. Four brands, three entities. The French SEO agency market is more concentrated than the number of names suggests, and two opinions you believe to be independent can come out of the same house.

A fifth domain, resoneo.com, returned HTTP 403 on all three attempts. That is not an observed block, it is a failed measurement: it appears nowhere in the figures.

What this measurement does not say

The list matters as much as the results.

• Sample of 27, no statistical test. Every gap reported is descriptive. A two-site difference moves a percentage by 7 points.
• The panel is not representative of the French market. It comes from an E-E-A-T audit plus 8 probed candidates, so it over-represents agencies already visible in SEO.
• Four entities are not independent SEO agencies and the survey labels them: a software vendor with a service arm, the French arm of a German agency, an acquisition agency where SEO is one line among several, an inbound agency. Removing them would change the figures.
• JavaScript rendering was not measured. No page was executed in a browser. A site whose promise lives in client-injected content is read mute: that is the case for one panel domain, 434 readable words across 2 pages.
• The truth of the published figures was not checked. The grid measures that a number is published and that it commits its author, not that it is accurate. No client case was cross-checked with the client.
• Price levels were not compared, only the presence of a quantified amount.
• Anything outside the public site is invisible: proposals, decks, social posts, unlinked campaign pages.
• One survey only, on 28 August 2026. No before and after comparison.

Three detectors were in fact corrected after the first pass, and the script carries a comment on each. The most instructive one: in French, a regular expression word boundary breaks on accents, so a pattern meant to catch "mes" fired in the middle of "mêmes" and filed a market statistic as an agency result.

Running the measurement on your three finalists

The full method takes half a day with no paid tool, and you do not need 27 sites: three will do.

1. Open the homepage and the service pages of each candidate, reading the source rather than the screen. What the HTML does not carry, AI engines do not read either.
2. Count the numbers in two columns. Theirs on one side, other people's on the other. The ratio shows up in ten minutes.
3. Look for a tool name inside the sentences about AI. Not in a logo carousel, not on a classic audit page.
4. Look for a method page and a pricing page. Their absence is not disqualifying, but it costs extra questions.
5. Put the 15 questions to them in the meeting and score out of 15. Compare against the panel: 9 is the median, 13 the highest observed.

For the definitions and the measurements that go with them, we have covered elsewhere what AI SEO actually means and how an AI citation rate is computed. Our own offer is described on the agentic SEO agency page, and our client cases are quantified, which is exactly the axis 21 sites in the panel fail to clear.