How Atyla measures
your visibility in AI.
An AI visibility figure is only worth something if you know how it was obtained. Here is what Atyla measures, how, and what the measurement does not say.
Questions asked every day
You choose the questions your prospects ask before choosing (30 on Starter, 60 on Growth). Atyla asks them every day to ChatGPT, Perplexity and Gemini and, on Growth, every 3 days to Claude, Mistral and Grok.
The same question comes back day after day, so there are several draws per question and per AI, not a single snapshot, which would prove nothing: asked twice, the same question does not get the same answer.
A neutral, localized instruction
AI engines queried through their API receive a single instruction: the user’s country and language (for example “the user is in France and writes in French”). It never asks them to cite brands, a given number of brands, or links. Web search is localized to the country when the API allows it.
For Gemini, Atyla reads the answer from the consumer app, the one your prospects see, through the provider DataForSEO, not only the API, which can answer differently.
This protocol has been in place since 7 October 2026. The previous one asked the AI to name several tools and give links, which inflated rates. Charts mark the date of the change: a movement that spans it should not be read as progress or decline.
Real, time-stamped answers you can read
Every figure opens on the answers behind it: the answer text, the AI, the date and time, and the cited links. You can check for yourself, and share proof with your leadership through a link readable without an account.


Named or linked
Atyla distinguishes two ways of being present in an answer: the brand is named in the text, or its site is cited as a source link. They are not worth the same and are not improved the same way: you can be named without being a source, and the other way round.
With or without your brand name
A question that contains your brand name (“is your brand reliable?”) almost always gets an answer that cites it. Mixing it with the others would inflate the figure. Atyla counts questions with and without the brand name separately: only the latter tell you whether an AI recommends you to someone who does not know you yet.
A margin of error, because a single figure lies
An AI answer is a draw. “34%” without a margin does not tell you whether next week’s 31% is a decline or noise. Atyla therefore shows every rate with its 95% confidence interval.
The interval is computed by question × AI clusters: the same question asked every day to the same AI gives similar answers; they are not independent draws. Treating them as such would give a margin that is too narrow, hence false precision.
- Below 30 answers over the period, a rate is not shown as a figure.
- A rise or a fall is only reported when it exceeds the margin; otherwise Atyla says “within the margin”.
Competitors, without public bodies
The ranking compares your brand with the other brands present in the answers. Public and reference bodies (statistics offices, regulators such as the AMF or the CNIL, public health insurance, standards bodies, encyclopedias) are excluded: an AI citing an authority is not recommending a competitor. They remain visible among cited sources.
The effect of each action, before and after
When an action is marked “Done” (a page changed, a reply posted off-site), Atyla compares the AI answers before and after, on the questions it targets. A first verdict comes 7 days later, a firmer one after 2 to 4 weeks, once enough answers have been collected.
If the gap stays within the margin, Atyla says so. If there are not yet enough answers to conclude, Atyla says that too, rather than showing a gain.


What Atyla does not do
- No invented “prompt volume”. No AI publishes how often a question is asked; such a figure can only be a guess, and we prefer not to show it.
- No guarantee of being cited. AI engines choose their answers and sources; Atyla measures, explains and prioritizes, but does not promise citations.
- No rise reported for noise: as long as a gap stays within the margin, it is not presented as a result.
- No measurement of what AI does not show: Atyla reads answers, not the internals of the models.
See the method applied to your brand, with the real answers.
Card required to start the trial, nothing is charged before day 8, cancel in one click.