Grounded answer

A large language model can answer a question in two very different ways. It can answer from memory — the text it absorbed while being trained, which may be months out of date and has no pages standing behind it — or it can search the web at the moment it is asked and answer from what it just found. The second kind is a grounded answer, and it is the only kind that can be treated as evidence of what a buyer sees today.

A grounded answer carries its sources: the engine hands back the URLs of the pages it read. Those links are kept with the answer, so the reading can be checked rather than trusted. Links an engine merely types into its prose do not count — only sources the engine returns as sources.

AgenticSEO measures grounded answers only. ChatGPT is asked after a live web search and may answer only from those results; Gemini is read from the real Gemini page, the way a person would see it; Google AI Overviews, where a client has its own search access set up, is read from Google's own answer together with the pages it cites. Engines that answer from memory are not asked at all.

When no engine produced a grounded answer for a buyer question, that question reads "not measured yet". It is never counted as a miss, never lowers a visibility score, and is asked again by the next audit. That is why a score is always stated as being out of the questions that were actually answered live.