How we measure AI visibility

The unit we count: the cited question

A question counts when at least one measured engine returns the client’s own domain among the sources of its answer. Not when it is mentioned in the prose, not when it might be implied — when the domain is in the citation list, where a reader can see it.

Per-engine answer counts are kept in the report as diagnostics: they show which engine is blind to you, which is useful for deciding what to fix first. They are not the headline number, because a business does not care whether two engines or four noticed it on the same question — it cares that the question now points at them.

The question set

A full audit uses 15 questions in the buyer’s own language, plus the locale, chosen on the audit date and stored with the report. Fifteen questions on four engines is 60 answers.

The free mini audit uses 8 questions on 3 engines. It is a different denominator, and it never carries forward: a sprint baseline is always re-measured on the full 15-question set. Two denominators for one business is expected and stated on the report rather than left to be discovered.

The engines, and the ones we refuse to score

Measured: ChatGPT, Gemini, Perplexity and Google’s AI Overview. These four return an answer and attach sources we can read and count.

• Grok is not measured. We found no text-answer endpoint we could query reliably, and a number we cannot take is a number we do not print.
• Claude is not part of any guarantee. It appears on paid sprints as a diagnostic column only. On the one measured call it returned, it named the client first in the answer and cited someone else’s page in the sources — a real and useful finding, but not a countable one.

What is recorded for every answer

• The answer text, as returned.
• Every source URL it cited.
• Whether the client’s domain is one of them.
• Which domains were cited instead, which is what turns “you are not mentioned” into an instruction.
• Whether the engine declined to answer at all.

The day-60 re-run

The comparison is only defensible if both runs are taken the same way, so:

• The 15 questions and the locale are frozen on the audit date and are the only inputs at day 60. No question is edited afterwards, in either direction.
• An engine that errors is retried up to three times. If it still errors it is dropped from both the baseline and the re-run, and the target is recomputed on the reduced set. A flaky endpoint must never be able to fail a client, or to rescue us.
• If fewer than two engines return at day 60, the comparison is void. We re-measure within seven days instead of arguing about it.
• The same provider is pinned for both runs. No switching suppliers between the two measurements.

The sprint guarantee, in arithmetic

The Fix Sprint guarantees the count at least doubles. The target is min(15, max(3, 2 × baseline)):

At a baseline of 8 the formula would already ask for 15 of 15, which is a perfect score and a promise nobody honest can keep. So the Sprint is not offered to a business already cited in 8 of 15: the audit says so plainly and the conversation becomes a retainer or nothing. A doubling rule is only honest for a business that is currently close to invisible.

What we do not claim

• No traffic numbers. We never see the client’s analytics, so we could not defend a lift percentage. Any agency quoting one is quoting something it cannot measure.
• No first-place promise. We promise a change in a count we take ourselves, twice, the same way.
• No score for an engine we cannot query. See Grok, above.
• No client-side setup. Intake is a URL. No analytics tag, no Search Console, no account access.