How we measure AI visibility

Seven surfaces, counted separately

ChatGPT, Google AI Mode, Google AI Overviews, Microsoft Copilot, Perplexity, Gemini and Claude. AI Mode and AI Overviews are counted separately because they are different products returning different answers from different sources. Search Console reports them together, and we think that hides more than it saves.

Perplexity and Claude are queried through their official APIs; Gemini and the other consumer surfaces come through licensed data providers, because we will not scrape a consumer interface ourselves and we will not pretend a developer API is the same thing a buyer sees. Claude is sampled three times per prompt rather than seven, at reduced reasoning effort with at most three web searches per answer — a cost decision, stated here because it changes what that row measures. Those providers collect answers from the public consumer products under their own terms; we are not affiliated with any engine, and an engine can block collection at any time — if a surface goes dark mid-audit, the report says so rather than filling the gap.

Twenty to thirty, human-edited before anything runs

Each prompt is a base term plus qualifiers drawn from audience, evaluation criteria, constraints and comparisons — the shapes real buyers type. They are grouped into three to five topics and capped there: spread thinner and a topic average becomes an average of noise.

No brand names appear in the prompts. Asking “is [brand] good for X” measures how politely a model answers a leading question. Asking “best X for Y” measures whether the brand gets recommended at all.

Seven runs, every prompt, every engine

Seven is not arbitrary. Published work on AI-answer stability finds day-to-day overlap in cited sources of roughly 0.34–0.42 and recommends at least seven runs before citation-level results settle. A standard audit is therefore 30 prompts × 7 engines × 7 samples = 1,470 answers, every one stored verbatim.

A clustered bootstrap, and where bands are allowed to appear

We report a percentage with a 95% interval. The interval comes from a clustered bootstrap that resamples prompts, not answers — seven answers to one prompt are correlated, and a plain binomial interval over 1,470 correlated draws would understate the true uncertainty by a wide margin.

Bands are reported at engine and topic level only. A per-prompt percentage from seven samples is not a measurement and we do not print one.

Where a composite across engines is shown, its weights are published and derived from measured referral share. Two of Google’s AI surfaces produce few attributable referrals, so their weight is a stated modelling allowance rather than an observation — we would rather tell you that than let it hide inside a single number.

Reading the brands out of an answer

A model extracts the brands, the order, the sentiment and the cited URLs from each stored answer at temperature zero. It is given the answer text and nothing else — never its own knowledge of the category, which is how tools end up confirming mentions that were never there.

Brand-name variants are folded by an exact-match pass over a maintained alias list, never by fuzzy matching. Anything the pass cannot resolve is flagged for a human and blocks the report until it is cleared. A matcher that guesses invents mentions, and undercounting a client’s own brand is the worst direction for a paid audit to be wrong in.

What we do not do

• We do not scrape consumer interfaces ourselves.
• We do not publish prompt search volumes as if they were measured. Where volume appears it is labelled as a Google keyword proxy, or it is absent.
• We do not report a single “rank”.
• We do not offer rank tracking, backlinks or site crawls. This tool does one thing.