Measuring AI answers
What AI answers, we measure.
When a customer asks ChatGPT, Claude or Gemini, the answer decides who they will choose. We measure these answers, we study how they are formed, and we built a way to act on what they find.
Why
AI has become the intermediary
It no longer shows a list of links: it builds an answer, and picks the companies it deems relevant. Those it does not cite do not exist in that answer.
The answer is built
“The answer no longer comes from a simple list of results. It is built. And that is where a profound change begins.”
Economic invisibility
“A company can fully exist in its market while being absent from AI answers. Not because it is bad, nor because it lacks customers: simply because the AI does not have enough usable information to recommend it.”
Silence cannot be seen
“False information can be checked. Missing information is different: you first have to know you should look for it.”
mAIr Observatory, 2026.
Our approach
Making the black box readable
A prism splits opaque light into a readable spectrum. We do the same with AI: answers that look arbitrary become signals that can be counted, compared and tracked over time. We do not control what happens inside the models; we control the conditions of the measurement.
- 01
Measure
Each question is asked twenty times, to each AI provider separately, to capture the real distribution of answers rather than a single case.
- 02
Understand
What the AI knows from memory and what it finds by searching the web are measured apart, never mixed. The gap between the two is precisely the information.
- 03
Count without AI
Presence, rank, share of voice, criteria: everything is computed by code, from the raw answers. The AI answers; it never scores.
- 04
Date and seal
Each measurement is frozen at the moment it is taken, then sealed so that no one can alter it. It can be checked.
What our measurements showed
Six findings from the Observatory
During the summer of 2026, we published our measurements in the mAIr Observatory. Each article separated what was demonstrated from what remained a hypothesis. Here is what stands out.
AI concentrates
Six brands
present in more than 8 answers out of 10, in a market where the AI cited 76 different competitors.
mAIr Observatory · 14 June 2026
The answer depends on the path
7.5% → 77.5%
Performance as a criterion for choosing a gift, between forty answers from memory and forty with web search. Same question, other priorities.
mAIr Observatory · 17 July 2026
Optimising is not measuring
0%
presence for a website rated A by an AI optimisation tool, with and without web search, while 57 other players were cited.
mAIr Observatory · 17 August 2026
Accessible is not consulted
1 measurement in 5
where the studied website appears among the sources: 100 citations out of 2,336. The AI does search the web; it does not necessarily search on your site.
mAIr Observatory · 24 August 2026
AI builds its own market
98 players
cited in 120 answers from several engines: the AI decides who is compared with whom, and on which criteria.
mAIr Observatory · 3 August 2026
Even us
0%
Our own brand was absent from all 40 answers to the question that described exactly what we do.
mAIr Observatory · 27 July 2026
A measured case
A brand the AI did not know
Asked from memory, three AI providers never cited it. With web search, one cited it every time, another six times out of ten, the third still not at all. The AI knew nothing about it: it was only finding it, or not. And 134 competitors filled the space left empty.
Real measurements, anonymised: artisan mills industry.
| from memory | with the web | |
|---|---|---|
| OpenAI | 0 % | 0 % |
| Anthropic | 0 % | 100 % |
| 0 % | 60 % |
From measuring to acting
Why we created FaireDuBruit
“An AI never decides better than the information it is given.”mAIr Observatory · 17 July 2026
Measurement after measurement, the same finding came back: what influences the models is what circulates on the web and what they hold in memory. The absent companies were not absent for lack of quality, but for lack of information where the AI looks.
Measuring tells you where you stand. It does not change what the AI finds. From this we built FaireDuBruit: spreading the company where AI assistants look, then recording their crawlers’ visits to the pages.
Act
Your website alone is not enough.
When a customer asks ChatGPT, Claude or Gemini who to choose, the AI answers with names. If yours is not among them, you are not part of their choice.
FaireDuBruit spreads your company where the AI searches.
Discover FaireDuBruit →of AI citations go to the brand’s own website
The other 86%: press, encyclopedias, institutions, other companies’ content.
Muck Rack, “What is AI Reading?”, May 2026: 25+ million links cited by ChatGPT, Claude and Gemini across 17 industries; 13.7% of citations go to content owned by the company the question is about.
Our principles
What does not change
- Facts, not opinionsWe show what is measured, with its method. What is demonstrated is kept apart from what remains a hypothesis.
- No promisesNo one can guarantee a place in an AI answer. At best, you raise the odds — and you measure it.
- The decision stays yoursWe provide the facts and the means to act. The strategy belongs to you.
- DiscretionWe never disclose named results of our customers or of their competitors.
Our goal
That no company is left out of an answer simply because the AI lacked information about it.
Contact
Let’s talk
For any question about our approach or FaireDuBruit:
Legal notice
- Publisher
- Lirenprism, Société par actions simplifiée, share capital 2 000,00 € — Chemin des Étangs No 1, 24120 Beauregard-de-Terrasson — 106 694 433 R.C.S. Périgueux — EU VAT number FR92106694433
- Publication director
- Émilie Charlet Santos, President
- Hosting
- OVH SAS — 2 rue Kellermann, 59100 Roubaix, France — RCS Lille Métropole 424 761 419 — phone 1007 (+33 9 72 10 10 07)
- Contact
- contact@lirenprism.com
- Personal data
- This page sets no cookies and uses no audience measurement tool. The server keeps technical logs (IP address, requested page, date) for security purposes.
- Emails
- Lirenprism no longer sends emails. Unsubscribe links in past messages lead to the unsubscribe page.