How we measure
AI assistant answers vary between users, sessions and days. Any tool that hands you an exact number without explaining how it got there is telling you half the story. Here is how we do it:
1. Buyer prompts, per product
For each product we generate up to 25 real recommendation, comparison or quote questions in its category, country and language — the ones a buyer would ask ChatGPT or Perplexity. They are editable: you know how your customer asks.
We measure what your customer sees, not the most powerful model
Each channel is queried with the model its app serves to consumers by default — not the most advanced one in the API. ChatGPT is measured through the official alias that automatically tracks its default model; Gemini and Claude, with the tier of their apps. When a provider changes its default model, we update the measurement and note it: a jump in your trend may reflect that change, not your content. Queries run with no history and no personalization — the reproducible baseline of a new user. Models last updated: July 21, 2026.
Google's channels are not an ordinary chat — and we measure them for what they are
For the Google channels (AI Overviews and AI Mode) your buyer question is queried as the search that would trigger it, anchored to your country, and we analyze the AI answer Google builds (with its cited sources). Two honest caveats: if Google shows no AI answer for that search, we count it as "not visible" — we don't hide it; and since the same search returns almost the same thing on the same day, these channels use one sample per measurement instead of three (repeating it adds no real variation and would narrow the interval artificially). And a third: Google does not serve results «from Latin America» — its search is always from a specific country. If your market is regional, we measure these channels in several representative countries and average them (Latin America in Spanish: Mexico, Colombia and Argentina). The country-by-country breakdown is in your dashboard, and in these channels the band reflects the spread between them.
2. Several samples per question
The same question runs several times per model (2 samples on a weekly cadence) or is averaged over a 7-day rolling window (daily cadence). A single answer is an anecdote; several samples are a measurement.
3. Share of Voice with a confidence interval
SoV = mentions of your product ÷ (mentions of your product + mentions of your competitors) across every sample in the window. We always show it with two measures of uncertainty: a 95% confidence interval (Wilson interval computed over every mention in the window) and the observed range, min–max across samples and models. If the range is wide, the AI is undecided in your category — and that is information too.
4. Trends, not absolutes
Today's exact number matters less than the direction: are you going up or down after landing that mention in the comparison site the AI cites? That is why we annotate completed actions on the timeline, to see the before and after of each one.
5. Mention ≠ citation
We distinguish a product mention (the AI names your product) from a source citation (the AI used a site as a source). Tracking both tells you whether the AI trusts your content and whether it also recommends you — and where you are missing. In the dashboard these ideas boil down to three metrics next to SoV: SoR (in what % of answers you show up among the recommended), Citation Share (of the answers that cite sources, the % that cite your domain) and SoT (in how many of the measured questions you show up at least once).
6. Real, editable buyer questions
The 25 prompts for each product are generated by an AI for your category and market, fed with real Google Autocomplete searches from your country — the language and intent people actually use. Each prompt carries a keyword and, when there is data, a search volume signal on a 1–5 scale. To be clear: Google search volume is a proxy — the real volume of questions asked in AI assistants is not public in any tool. And the prompts are yours: you can edit them, turn them off or add the ones your customers actually ask.
Questions about the methodology? Write to us. Transparency is the only honest way to measure something as variable as an AI answer.