How Recoia measures visibility in AI assistants

1. The engines we query

We measure ChatGPT, Gemini, Perplexity and Google AI Overviews. Answers are obtained through data providers and official APIs — never by automating the consumer interfaces. Each answer is stored with the sources it cites, its date, the engine, the language and the location.

2. The questions we ask

Questions are not keywords: they are realistic requests, such as “Which plumber would you recommend in Brussels for an urgent leak?”. They are generated from the website being analysed by combining its services, areas, target customers and intents:

• Discovery — “who can help me with…”
• Comparison — “what is the difference between…”, “best…”
• Decision — “how much does… cost”, “who should I contact in…”
• Brand — “what do you think of [business]?”

Questions are asked in the language of the market being measured: French for France and French-speaking Belgium (Dutch for Flanders is planned). Each question is weighted by real demand (search volumes), so that visibility reflects what customers are actually looking for.

3. Several runs, always

In every measurement, each question is asked 2 times to each engine. A site’s indicators cover dozens or hundreds of questions, so hundreds of answers: we never display a score calculated from a single answer.

Tracking runs every 2 weeks by default, weekly on request, or only when you start a measurement. The number of questions tracked is suggested from the size of the site (offers × places) and can be adjusted.

4. Rates with their margin of error

For every indicator (mention rate, citation rate with a link, share of voice), we calculate a 95% Wilson confidence interval. It shows the range in which the true value is very likely to lie. The more answers we have, the narrower the range:

In practice: a drop alert is only sent when the new measurement falls outside the interval of the previous one. No false alarms caused by chance.

5. What we extract from each answer

• the brands and businesses mentioned, and their position in the answer;
• whether the website is cited with a link or only mentioned;
• the sources, classified by type: brand website, directory, comparison site, media, forum, review platform;
• the sentiment (positive, neutral, negative).

6. Pooled panels

The same normalised question, for the same engine, language and location, is run only once per period, then reused for every website concerned. That is what keeps audits affordable — and it is statistically cleaner too: everyone is compared on the same answers.

7. Transparency

Every raw answer can be read and filtered in the app. If a figure surprises you, you can always read what the AI assistant actually said.

To go further: our articles and the page What is Recoia?.