Method & limits
How we measure what AIs say
Opinion Radar builds questions that represent the requests made to AI assistants, has a human approve them, then analyzes the responses systematically. Here is the protocol, and its limitations.
Four steps, always the same
Representative situations
We start from moments when someone asks an AI for advice: ‘which tool should I choose?’, ‘what should I think about…?’ Opinion Radar suggests them and a human approves them before measurement.
Questions asked without prompting the answer
In a visibility measurement, questions do not name the tracked brand. If an assistant mentions it, the mention is spontaneous; observatories label direct questions separately.
A panel of models
Each question is asked to 6 AI families (ChatGPT, Claude, Gemini, Le Chat, DeepSeek, Grok), several times: answers vary from one draw to the next.
A systematic reading
Each answer is analyzed: who is mentioned, with which argument, what tone, what caveats. Report figures are aggregates of these readings.
Three simple metrics
Share of voice
The share of answers where an entity is mentioned spontaneously. 63% = mentioned in 63% of the answers measured across the sector.
Tone
Favorable, neutral or critical: the framing that comes with each mention, with verbatims so you can judge for yourself.
Dodging
The share of questions a model refuses to answer squarely: often the most telling figure on sensitive topics.
What this measurement is not
- A ranking of the actual quality of products or people.
- A feature or price comparison.
- An opinion poll: we measure the behavior of AI models, not what people think.
- A stable truth: models evolve, every measurement is dated.
Transparency
Every public report shows its measurement date, its number of answers and the models surveyed. Rates computed on few answers are flagged as indicative (“n=”).
The method, applied in public
Our observatories are real, dated, freely accessible measurements: the best way to judge the method.