How RadarIA uses AI
RadarIA measures what AIs answer. It is only fair to tell you where an AI is at work inside RadarIA, where it decides nothing, and what is re-read before it reaches your report. Everything below describes the product as it works today.
8 AIs queried every day, with reproducible questions
Every day, RadarIA asks the same questions to ChatGPT, Claude, Gemini, Perplexity, DeepSeek, Mistral, Google AI Overviews and Google AI Mode, through their official APIs (for Google, which has no API for its AI answers: the results page, collected by DataForSEO). The questions (15 per brand, 30 with the 30-question option) are written the way a buyer would type them; your agency reads them, edits them and approves them before they are asked. It is the questions, not the answers, that are fixed: the protocol is the same from one measurement to the next, and an AI that is down on a given day is marked as absent rather than counted as zero.
Cited brands are picked out by a model, then checked
To know who is cited in an answer, a language model extracts the brands, their rank and the tone. Two checks that use no model run behind it. The first removes any brand the extraction claims to have seen that is not in the text of the answer. The second re-reads the raw text and counts omissions: a name written in the answer that the extraction missed. The omission rate is shown next to the score. Nothing is corrected silently: a figure that has already been sent is not rewritten without a human decision.
“Why a competitor comes first”: an automatic analysis, to be reviewed
When an answer does not cite your client, an AI re-reads the answer, the brands found and the sites consulted, then proposes a verdict and a lead. The analysis is shown with the note “Automated analysis, to be reviewed by your agency before any action.” Safeguards run before anything is displayed: a truncated answer, a client who is actually present under another name, a lead that names a site or a brand absent from the answer, or a proof that cannot be found in the answer all stop the analysis from appearing. It never changes a score. We do not present it as reliable until a wider audit has concluded: it is an aid to review, not a verdict.
Reports are built from the measurements
The figures in the Monday report (score, change, AI by AI, competitors cited) are computed from the answers measured; they are not written by an AI. The expert’s comment that goes with them is generated automatically by a language model, from the report’s figures only, with an instruction to invent no figure or outside cause and to promise nothing. If it cannot be produced, the report goes out without a comment. Nothing is published on your agency’s behalf: the corrective texts are first drafts to be reviewed.
Your data trains no model
RadarIA does not train any model. To the AI providers we send only the name of the brand analysed, its sector and the question asked — not your email address, not your agency’s identity. Scans run through the professional interfaces of OpenAI, Anthropic, Google and Perplexity, whose terms exclude training their models on the data sent. RadarIA neither sells nor passes on this data.
Where the data is hosted
The database is hosted in the European Union (Supabase, Paris), as are email delivery (Resend) and the encrypted backups (Backblaze, EU). The site and function execution (Vercel) and the AI providers are companies established in the United States; those transfers rely on the European Commission’s standard contractual clauses. So this is not “everything in Europe”: the full list of processors is in the data processing agreement (GDPR, article 28), available on request.
DeepSeek (China, Hangzhou), when it is among the AIs queried, receives only the protocol’s questions and brand names — no account data.
Mistral AI (France, Paris), when it is among the AIs queried, is established in the European Union: no transfer outside the EU for that call. It receives only the protocol’s questions and brand names — no account data.
Google AI Overviews and AI Mode, when they are among the AIs measured, are not queried through a Google API (there is none for these answers): DataForSEO (Estonia, Tallinn), established in the European Union, collects the Google results page for the question, the brand’s country and language. It receives only the protocol questions, no account data.
European AI regulation
RadarIA measures how brands appear in AI answers; it makes no decision about individuals. As we read Regulation (EU) 2024/1689 on artificial intelligence, it is not a high-risk AI system. As for transparency, that is the purpose of this page: to say where AI is at work in RadarIA (extraction, “why” analysis, expert comment) and what is to be reviewed.
Going further
Last updated: 15 October 2026