How we measure
Draft copy — product content, not final. Written to be true on 2026-08-26; every
sentence here passes site/check_claims.py today. Needs a French translation before
publication (see site/README.md).
Most tools that claim to measure “AI visibility” won’t tell you how they got the number. We will, including the parts that make our numbers look smaller.
What we actually do
We write the questions your buyers ask before they buy — in French, the way they’d type them — and we ask an AI assistant each one. Then we read every answer and record which companies it named, in what order, what it said about them, and which web pages it drew on.
Nothing about that is magic, and you could do it yourself. What takes the time is doing it the same way twice, so that this month’s number and next month’s number mean the same thing.
What “you appear in 12% of answers” actually means
12% is an example here, not a measurement — your own number comes from your own batch.
It means: of the answers we collected in one fixed batch, 12% named you somewhere in the body text.
Three things worth knowing about that number before you rely on it:
Half of all answers name two products or fewer. Not ten. The average is a little higher — 2.82 across 132 answers — but that average is pulled up by a minority that list eight or more, and 4 answers named no product at all. The median is 3. So if your category has ten serious players, nobody is going to appear in most answers — the ceiling is low for everyone. That’s why we lead with the ratio between you and your competitors rather than the raw percentage. “Named four times more often than you” is the fact that matters; “18% versus 5%” is the same fact rendered less useful.
Questions that name you don’t count toward that number. If we ask “Is [your brand] reliable for a small company?”, the answer mentions you. Every time, for every brand we’ve tested. Including those would inflate your score for a reason that has nothing to do with your visibility, so we report them separately, as a different measurement.
Sample size sits next to every percentage. With 40 questions, one answer is 2.5 points. We won’t tell you a two-point move is a trend.
How we check whether the models are wrong about you
This is the part we care most about, and it’s the part where it’s easiest to cheat.
We fetch your own website — around fifteen pages, respecting your robots.txt — and
compare what the model said about you against what your site actually says. Every finding
comes with two quotes: what the model said, word for word, and what your page says, word
for word.
The shape of a finding, with an invented company — Exemplia does not exist:
ChatGPT told a buyer: “Exemplia propose une formule gratuite pour les TPE.” Exemplia’s pricing page says: “Formule Starter — 49 € HT / mois”
Neither of those is written by us or summarised by us. Both are exact text, located by literal search, with the source page and the date we fetched it. If we can’t produce both quotes, the finding doesn’t go in the report — we’d rather show you less than show you something you can’t check.
We report three outcomes, not one:
- Contradicted — the model says something your own site disagrees with.
- Not addressed — the model says something about you that appears nowhere on your site. That doesn’t make it false. It means it came from somewhere else, and you may not know where.
- Supported — your site backs it up. We report this too, because a section that only ever finds problems is fishing.
What we don’t do
We don’t cloak. We will never advise you to show AI crawlers something different from what you show people. It’s a deindexation risk for you and a line we won’t cross.
We ignore nothing about your robots.txt. We read it, we obey it, and if it blocks
something we needed, the report says “couldn’t verify” instead of quietly working around
it. We also tell you which AI crawlers your own robots.txt blocks — occasionally that’s
the entire explanation for a bad number.
We don’t use search-volume data and call it AI data. Some vendors take Google keyword volumes and present them as what people ask assistants. They’re not the same thing.
We don’t invent competitors. The companies in your report are the ones the models actually named in answers we collected. That list often differs from the one you’d write yourself, and the difference is usually the interesting part.
Where this method is weak
An honest list, because you’ll find these anyway:
We query the API, not the consumer app. What we measure is close to what a person sees in ChatGPT, but it isn’t identical — different interface, different defaults.
One measurement is one moment. These models aren’t deterministic. Ask the same question twice and you can get different answers. A single snapshot tells you roughly where you stand; it can’t yet tell you that a small change since last month is real rather than noise.
Absence is harder to prove than presence. “You were never named” and “you were unlucky in a small sample” look the same from 40 questions.
We only see the pages we can fetch. Some sites block automated requests or render everything in JavaScript. When we can’t read a page, we say so rather than guessing.
Every number in your report is reproducible from stored data. If you want to see the raw answer behind any figure, ask — we keep all of them.