They are live answers, not a canned dataset. Rixe calls each engine's search-grounded endpoint — the retrieval-backed models that answer real users — once every 24 hours, and stores the response verbatim with its citations. What you see in the dashboard is the answer that engine produced that day.
The distinction that matters is retrieval. A plain model call with no web access answers from training data alone: it will happily name brands that no longer exist and miss the review that changed everything last month. Rixe only queries search-grounded surfaces — the retrieval-backed endpoints behind ChatGPT, Gemini, Perplexity, Google AI Overviews and the rest — because those are the paths that actually read the open web before answering, which is what makes the result comparable to what a user gets.
Answers also drift. The same prompt asked a week apart can name a different third brand, because the sources moved, the index refreshed, or the model was updated. Measuring daily is what turns that drift into a trend line instead of noise — and it is why a fall in visibility can usually be traced to a specific week and a specific source.
Because the response is stored verbatim with its citations, anything in the dashboard can be quoted directly into a report or shown to a client without paraphrase.
Retrieval is the distinction that matters
A plain model call with no web access answers from training data alone. It will happily name brands that no longer exist, quote a price that changed two years ago and describe a product that shipped a rewrite since. Measuring that tells you what a model remembers, not what a buyer is being told today.
A search-grounded call is a different object. The engine goes to the live web, runs its own searches, reads what it finds and writes an answer from that — which is what a real user gets, and the only thing worth counting. Every figure here comes from that second kind of call.
What one run actually does
A run is one prompt, sent to one engine, in one language, from one market, on one day. The answer that comes back is stored whole, then read for the brand, its position, the tone of the sentence around it, the sources cited and the competitors named. The stored text is what makes the number auditable later.
Runs are spread across the day rather than fired in a burst, because engines rate-limit and because an answer at nine in the morning is not guaranteed to match an answer at six in the evening. The schedule is part of the measurement, not an implementation detail.
What is deliberately not done
No panel data, no scraped third-party dataset and no modelled estimate stands in for a run. If an engine was unreachable, that day is marked incomplete; if a prompt has too few runs, no score is published for it. A gap that is visible is worth more than a gap that has been filled in.
The same rule governs what an answer is allowed to claim about a brand. A sentence an engine produced is quoted as that engine's output on that date, never rewritten into a cleaner version; if the classification of its tone looks wrong, it can be flagged, and the flag travels with the evidence it came from.
