Track what an AI answer can prove.

Whether an assistant mentions your business decomposes into three questions. They are strictly ordered — no amount of work at layer three compensates for a failure at layer one.

• Access. Do the public directives tested by Revvye allow the named agents? The live tool evaluates robots.txt, homepage meta directives, and response headers; agent-specific firewall behavior requires separate inspection.
• Legibility. Once fetched, do your pages state anything an assistant can extract and repeat — a business type, a location, services, prices, hours — in a form that survives being read by a machine? Largely a structured-data and markup question.
• Mention. For the questions your buyers actually ask, does the assistant name you? This depends on the first two, on how much of the wider web corroborates what your site claims, and on the model's own behaviour, which changes without notice.

A rank tracker works because a query returns a stable, ordered list that is roughly the same for everyone. AI answers have none of those properties. Two people asking the same question in the same minute can get different answers with different citations. The same question asked twice can vary. There is no position to occupy.

What can be measured is presence across a defined set of questions, sampled over time. You choose the questions that matter to your business, ask them repeatedly, and record whether you appear, who appears instead, and whether the assistant links to a source. That is a different instrument from a rank tracker and it should be read differently — as a trend across many samples, not as a number.

A useful prompt set is small, specific, and drawn from how buyers actually speak. Four families cover most of it:

• Branded — questions that name you directly. If an assistant cannot answer 'what does [your business] do', nothing else will work.
• Category — the service plus the qualifier a buyer would use, with no brand attached. This is where you either appear or a competitor does.
• Local — the same category questions with a place attached, which is where most service businesses actually compete.
• Comparison — you against a named alternative. These surface how the assistant characterises you, which is often more revealing than whether you are listed.

A small, documented question set tracked consistently is easier to compare than a sprawling set that changes between runs. The point is to inspect change, not imply universal coverage.

The public directive and page-structure layers are inspectable from outside, within the limits of a successful fetch. The free tools report the exact surfaces they retrieved and expose limited or failed states rather than inferring a clean result.

Layer three is sampled rather than ranked. An assistant's answer is a generated response influenced by sources and system behavior an external tool does not control. Revvye can calculate rates from an evaluated prompt set, but the denominator and run evidence must remain visible. The rate describes that sample, not the whole market.