How we measure, and what our figures do not say

• We start from what the client sells: the services and products read on their site, and the Google searches they already rank for.
• A model filters out what is not an offer — materials used, off-topic subjects, searches for the company's own name — and identifies the core business.
• For each topic we estimate demand in answer engines. Whatever has proven demand — at least 10 estimated searches a month, or an AI answer Google already shows — comes first. The core business is still measured when the estimate misses it; it is then flagged "demand not measurable".
• The question is sent the way a customer types it ("loft insulation Nantes"), with no wording that would push the engine to name brands. For a local market we add the city: that is what makes engines name companies.
• The agency sees the topics left out and why, and can add or remove prompts.

Each engine is queried programmatically, with web search on when it offers it, in the client's country and language. Measurements are scheduled: every figure carries its collection date.

The model queried programmatically is not always the one in the consumer app, and it does not know a user's history: this is a measurement comparable from week to week, not the exact answer each person will see.

• Its name, or one of its aliases, appears in the answer as a whole word;
• or its domain name appears in the answer;
• or a page of its site is among the sources the engine cites.
• We also record the position of the first mention, the competitors named instead, the sources cited and the tone of the answer.

• The same question does not always get the same answer. Every percentage is therefore shown with the number of answers it rests on (n) and its 95% margin of error (Wilson interval) — for example "50% (±28 pts, n = 8)".
• To narrow that margin on the prompts that matter most, the same question can be asked up to 5 times per measurement.
• A change is only reported when it rests on at least 8 answers and exceeds what the measurement can tell apart. Otherwise we write "no change shown" rather than a rise or a fall.

• No provider sells the actual number of questions asked to ChatGPT in France. The demand we show is a DataForSEO estimate, computed from Google search signals, and it is not broken down by engine.
• It is used to choose and order what we measure. It is never presented as a guaranteed audience.

The audit checks 45 criteria, in four families.

• It does not predict the exact answer each user will get: history, signed-in account and wording change answers.
• It does not measure traffic or leads: visits from AI are tracked in your analytics tool.
• It guarantees no citation: we measure and recommend; the engines choose their sources.
• Gemini, Mistral Vibe and Claude are measured less often than ChatGPT and Google AI Overviews: their figures can be a month old.
• The demand estimate comes from a model: it can be zero on a very real niche.