An honest AI visibility measurement is one somebody else could reproduce: a question set stated in advance, conditions controlled and recorded, repeat runs on every question, results reported per surface rather than blended, and the raw responses kept. This chapter covers what that takes, and why the result is a fraction rather than a percentage.
Why can't you trust a single AI-visibility score?
Most AI-visibility numbers on offer are a single composite score whose derivation the buyer cannot inspect. That is not a small complaint. A number you cannot check is a number you have to trust, and trust is exactly what an agency should be earning rather than requesting.
A score also always moves in the flattering direction, because whoever built it chose what goes into it.
What does a defensible AI visibility measurement require?
A defensible measurement needs five things: a stated question set, controlled conditions, repeat runs, per-surface reporting, and kept evidence. Each of the five is set out below, and each is there for the same reason, which is that the number at the end has to be checkable by somebody who did not produce it.
A stated question set. Which questions, chosen how, and why those. Twelve real buying questions rather than whatever produced a good result.
Controlled conditions. Signed out, fresh session, no history, location set explicitly and recorded. Personalisation cannot be eliminated, only reduced, and the residual variance should be stated rather than hidden.
Repeat runs. Three per question per surface, because these systems are non-deterministic and one run is close to meaningless. A question scores on a majority, and splits get reported as unstable rather than rounded away.
Per-surface reporting. Never blended. A single figure averaging ChatGPT, AI Overviews and Perplexity hides the platform where you are absent, which is the most useful line in the report.
Kept evidence. Full-text responses, timestamped, per run, handed over. The number exists so it can be checked.
What is the difference between a mention, a recommendation and a citation?
Mention, recommendation and citation are three different outcomes with three different definitions, and they are counted separately rather than added together. The table below gives each one its definition. The paragraph after it says which of the three a headline number should be built on, and where the other two go.
| Outcome | What it is | Source |
|---|---|---|
| Mention | A mention is your name appearing. | Mindflow measurement protocol, published at /methodology/. Not a platform publication. |
| Recommendation | A recommendation is the response putting you forward as an option. | Mindflow measurement protocol, published at /methodology/. Not a platform publication. |
| Citation | A citation is your own domain named as a source. | Mindflow measurement protocol, published at /methodology/. Not a platform publication. |
Collapsing these is the commonest way an AI-visibility figure gets inflated, because mentions are far more frequent than recommendations. A headline number should count recommendations, since that is what a buying question is actually asking for, with the other two reported alongside and never folded in.
Why is the result reported as X/12 and not a percentage?
Twelve questions cannot honestly produce a percentage. Reporting 8/12 as 66.7% implies a precision the sample does not support, and the decimal is doing rhetorical work the data cannot back. The score therefore stays a fraction, with its denominator left where the reader can see it.
Our full protocol, question selection, intent validation, surfaces tested, personalisation controls, cadence, the exact calculation, repeat-run handling, evidence capture and known limitations, is published in full. Not summarised. Published, so you or another agency can reproduce the number.
Rankings are never guaranteed. Anything we could not trace to a primary source is absent from this page, not estimated. The audit runs the same six layers described on the pricing page, and Share of Answer is scored quarterly.
Read in order, or jump
Twelve chapters, in the order the work has to happen. The first four are the spine; the platform chapters at the end include two surfaces we deliberately do not score, and say why.
01 How AI search picks local businesses 02 Eligibility: can AI even see your site? 03 Robots.txt: block AI crawlers or allow them? 04 Measuring AI visibility honestly 05 Entity foundations for local businesses 06 Reviews as AI input 07 Citation surfaces: where AI engines look 08 Getting ChatGPT to recommend your business 09 Showing up in Perplexity 10 Showing up in Google AI Overviews 11 Showing up in Gemini 12 Copilot, and why we don't sample itSee where you actually stand
Everything measured in this guide, run against your business, free.
Get my free Visibility Check