How we measure

1. AI models we measure

We use two collection methods, shown next to each answer:

• Captured: the answer shown in the consumer product, collected from the product page by a data-collection provider.
• API simulation: a call to the provider's model with a web-search tool the model may or may not use. It is not the consumer product (model, instructions and sources can differ); these AI models are always labelled "API".

Location follows your prompts' region. For France, both Google AI models are queried from Paris; the AI Overview is also requested in French (fr-FR) and in its full version. ChatGPT is queried through a French connection without an account; by default it decides on its own whether to search the web. The Gemini API receives no location: this is a limit of that simulation.

2. Frequency and samples

Prompts are run roughly every 24 hours on every enabled AI model, within your plan's limits. AI answers vary from one run to the next, so we never draw a conclusion from a single answer. The main metrics (visibility, share of voice) show how many prompts they rest on and their margin of uncertainty.

In our reports, a per-prompt, per-AI-model result rests on at least 10 answers; below that it is marked "indicative". A day without answers (a provider outage, for example) shows as missing data, never as a zero.

3. Metric definitions

• Visibility (per AI model): share of analysed answers that mention your brand. For Google AI Overviews it is computed only on searches where an Overview was shown; the AI Overview answer rate (searches with an Overview ÷ searches) is reported separately.
• Overall visibility: plain average of the captured AI models (AI Overviews, ChatGPT, AI Mode) that have analysed answers in the period. API simulations are shown separately and never enter this figure. It covers "unbranded" prompts, which do not contain your brand name; prompts that name it are analysed in a "Reputation" block.
• Share of voice: your brand's mentions ÷ (your brand's mentions + mentions of the competitors you track). Each brand counts at most once per answer; the ratio is pooled per AI model, then averaged. It is shown only once you track at least one competitor; other brands cited are offered as suggestions.
• Sentiment: tone of the answers that mention your brand, and only those; "—" when none do.
• Rank: position by mention rate among the brands you track, computed on exactly the same answers for all of them. It is not a "position in AI".
• Changes: in percentage points (see section 4 for when they are coloured).
• Sources: links cited in the period's answers, with tracking parameters removed. A domain's citation rate is the share of answers that cite it.

4. Uncertainty and significance

Answers to the same prompt resemble each other from day to day: a prompt tracked for 30 days counts as one prompt. For the dashboard's main metrics, the 95% confidence interval is therefore computed at the prompt level (resampling prompts, a "bootstrap"). A change is coloured there only when both periods cover exactly the same prompt × AI model pairs, there are at least 10 prompts, and the 95% interval of the change excludes zero. Below 50 distinct prompts, these metrics are flagged as "exploratory".

For a single prompt on a single AI model, the interval is computed over its answers (Wilson interval). Source changes use a simpler rule and should be read as trends.

5. How answers are analysed

Each answer is read by a language model hosted on Azure OpenAI, which detects mentions of your brand, the other brands cited and the tone. Other brands' names are normalised (case, accents, legal forms) and matched to the competitors you track and their aliases. Raw answers can be consulted from each prompt's detail page, with their date, AI model and recorded model version.

6. What the AI searched

Before answering, an AI assistant often runs several web searches of its own, worded differently from your question ("query fan-out"). ChatGPT does not show these searches, so we observe them through OpenAI's API: the same family of model writes the searches and OpenAI's search runs them. They are close to, not a copy of, what ChatGPT searched for a given answer, and we label them that way.

• Each prompt is checked as soon as you track it (or reword it), again two weeks later, then every two weeks (every month on the Starter plan), or with every answer when you track the ChatGPT API.
• We show each search, how many checks ran it, the words the AI added to your question, and the pages its searches found: yours, your competitors', others'. Pages belong to a batch of searches, never to one wording.
• Google (AI Overviews, AI Mode, Gemini) is not covered: its API terms do not allow us to analyse its searches.

7. Known limits

• We measure your set of prompts under the conditions above, not the whole French audience.
• Personalised answers (signed-in accounts, history, memory) are not reproduced.
• API simulations can differ markedly from the consumer products; we show them separately.
• Not covered today: Perplexity, Microsoft Copilot, Mistral Le Chat and Meta AI.
• Captured answers depend on a third-party provider and on public product pages, which change without notice.
• Raw answers and their analyses are kept for 90 days (see our privacy policy).

A question about a measurement? Email us at support@deepfirst.ai.