What we measure, and what we don't.
The whole product is a measurement. This page is the instrument's label: how the questions are asked, what gets recorded, and where the method stops.
The questions
A Snapshot is 15 prompts across 3 to 5 verticals. We read your site and propose them; you edit and approve every one, so nothing runs on our assumptions. The set leans unbranded (what a buyer asks before they know who to call), with two branded prompts to check the engines know your company exists at all, and one comparison-framed prompt. Regulatory instruments only get tested in the jurisdiction they belong to. No cross-border shortcuts.
The engines
| Assistant | Model | Access |
|---|---|---|
| ChatGPT | GPT-5 | OpenAI API, web search on |
| Claude | Claude Sonnet 5 | Anthropic API, web search on |
| Gemini | Gemini 3.6 Flash | Google API, search grounding on |
| Perplexity | Sonar | Perplexity API |
Engines as of 14 September 2026. Every report states the model ids it used. Before each run we confirm each id still resolves, because a retired model fails a whole column silently, and silent failure is the one thing we won't allow.
Each prompt is asked once, in a fresh single-turn conversation, from your market's locale and location (London, GB unless you choose otherwise). Answers via a vendor's API sit close to, but not identical to, what the consumer app shows: same models, different retrieval, different personalisation. We're reporting a sample, not a census. Ask the same question tomorrow and you might get a different answer. That's the nature of the thing, not a flaw in how we counted it.
What gets recorded
Mentioned
Your brand (or an approved alias) is named in the prose of the answer.
Cited
One of your own domains appears in the answer's source list.
Rank
The position of your first citation within that source list.
Mention and citation are different channels, on different timescales. A brand can be named from training data and never cited, or cited by a retrieval engine while too new to be named from memory. We never blend them into one score. One number would hide the only difference you can actually act on.
Every match is checked
Short names collide; an alias can also be an unrelated product. Every matched mention is printed verbatim in the report and checked before it's counted. You can see exactly what we counted, and why.
The verdicts
For each vertical we also read what the assistants said: which vendors got recommended, and what each was credited with. Every extracted claim carries a verbatim quote from the frozen answer. A claim whose quote can't be found gets dropped, and the drop rate is reported. That reading gives each vertical one of three verdicts.
- Recognised, not cited — named in a meaningful share of answers but rarely cited: a page problem, not a brand problem.
- Entrenched — one rival appears in the majority of answers: the engines have a settled reply.
- Open — no vendor is named in more than about a quarter of answers: there's no settled reply to displace.
Thresholds are stated on every report next to the counts they were applied to, so you're never taking our word for where the line sits.
The free readiness check
The readiness check: this answers a different question entirely: can assistants read your site at all? It fetches your homepage, robots.txt and llms.txt once, identifying itself honestly as VisibilityToAI-ReadinessCheck and never impersonating a vendor's crawler. Six deterministic checks: crawler access by token, homepage response, visible text without JavaScript, structured data, llms.txt, indexing signals. Nothing is stored. A site that only serves content to crawler user-agents will show thin here, and that's a fair reflection of what a plain fetch actually receives.
What this cannot tell you
- It cannot tell you traffic; a citation is a source link, not a click.
- It cannot tell you why a model named you; training data and web sources aren't separable from the outside.
- It cannot promise stability; engines change, and that's exactly why Track exists.