How we measure AI citations

How a check runs

For each of your tracked keywords we query every engine and read what comes back: 5engines are sent one prompt with the engine's web search turned on, and Google AI Overviews and AI Mode are read from a real Google results page rather than answered by a model. A citation is recorded either as a matched source URL, or as an unlinked mention of your domain in the answer text. Your dashboard shows the two separately.

The prompt is not blind. For 4 of the 7engines we name your domain and ask the model to state whether it appeared. We do not take that answer on its own: a “cited” verdict is only recorded when we can corroborate it, either by matching a returned source URL to your domain or by finding your domain in the answer text. A verdict with neither is discarded. Perplexity is different, and stricter: we send the keyword alone and match the returned source list against your domain ourselves.

Google AI Overviews and AI Modeare asked nothing. They are surfaces on Google's own results pages, and for them we read the page: for each keyword we fetch a real Google results page and record whether an answer was present on it and whether your domain was among the sources it cited. Measured on US Google, English, desktop — a fixed combination, not your visitors' actual locations. Nothing is simulated and no model is asked to report on itself; a results page has no free-text answer of its own, so Google AI Overviews and AI Mode record URL-matched citations only. That makes three Google surfaces, three different measurements: Google Gemini answers our prompt through grounded retrieval, while Google AI Overviews and AI Mode are read from the results pages themselves — and each engine name is exactly one measurement, never another surface stored under a second name.

Because we ask once rather than repeatedly, a citation is a single observation at a single moment. We therefore do not report variance, confidence intervals, or “cited in 2 of 3 runs”. Some engines run more than one web search internally while answering that single prompt; that is the engine's decision, not a second sample by us.

Two things on this site are not tracked measurements. An Enterprise on-demand check runs one keyword you choose, so it is a spot check rather than a site-wide reading. The free citation checker on our tools pages runs three questions we generate about the domain, on Perplexity only, and is a preview.

What the numbers mean

The citation total on your dashboard is the running count of citations we have stored for the period you are viewing, from the engine checks described above. The count behind the 90-day guarantee is produced separately: it is read directly from the engine responses at each checkpoint, over the keyword set recorded when your baseline was taken, and it is not the same figure as the dashboard total for the same window. The guarantee terms, including how subscriptions that started before 1 August 2026 are measured, are on our terms page and are authoritative.

A refusal to measure is treated as a refusal, not as a zero. If a guarantee checkpoint cannot be completed across the full set of engines and keywords your baseline used, we do not write a checkpoint and the claim goes to a person rather than being decided automatically.

The position we show is the place your source appeared within that one answer's citation list. It is not a search ranking. Sentiment is inferred from the wording of the answer, not measured.

What we cannot see

• Google's panels beyond one fixed combination. We do measure Google AI Overviews and AI Mode— from real Google results pages, never from a model's self-report — but at one fixed combination only. Measured on US Google, English, desktop — a fixed combination, not your visitors' actual locations. A measurement location you set for a site does not apply to Google AI Overviews and AI Mode in this version, and what a logged-in person in another place sees on those panels, we do not capture. Neither engine is part of the 90-day guarantee basis: the guarantee keeps the frozen 5-engine basis set out in our terms. Google Gemini is a separate measurement through grounded retrieval, and its redirect-wrapped source URLs cannot be matched to your domain, so a Gemini citation is usually recorded as an unlinked mention rather than a matched URL.
• Microsoft Copilot. Microsoft publishes no citation API, so we never run a citation check against Copilot and it is excluded from the guarantee. Where Copilot sends a reader to your site, that visit appears in your analytics as referral traffic. It is not counted as a citation.
• Paid placements.We read each engine's own citation or grounding list. Those lists do not mark a paid placement as paid, and we do not attempt to separate them. As assistants introduce advertising, our numbers will not distinguish it from organic citation.
• What a logged-in person sees. Unless you set a measurement location for a site, we send no location, profile or memory — so those numbers are logged-out, unpersonalised numbers. Someone with assistant memory enabled, in a different country, may see a different answer to the same question. That is a real limit on any measurement of this kind, including ours.
• A negative, before August 2026. We now record every engine we asked and what it returned, so “this engine did not cite you” and “this engine was not asked” are distinguishable. That was not true of checks run before August 2026: those stored only the citations found, so for older data the two still look the same.
• Keywords beyond the checked set. A check covers up to 10 of your tracked keywords. Citations you earn for a keyword outside that set exist, and we did not look for them.