Build a visibility test you can explain.

Start with the reader or customer tasks you want to understand. Use real questions where available and label invented practice prompts as ideas. Keep branded questions separate from category questions: explicitly asking about a business makes its appearance mean something different from being named in a broader comparison.

Record the exact wording, purpose and prompt-set version. Choose a manageable collection that covers the tasks you can actually review. There is no universal number of prompts that turns a convenience sample into a representative survey. If you later add questions, preserve the old version so a changed mix is not mistaken for a visibility change.

Decide what counts as a mention, an owned reference and an answer you can evaluate. For this lesson, a mention means the organization is explicitly named in the saved answer; an owned citation means at least one displayed reference points to its website. A repeated name in one answer still counts once for the answer-level mention measure.

Record the product, mode, date, market/language context you can establish, prior conversation and collection method. If a field cannot be observed, say unknown. Keep source verification and factual accuracy as separate notes; a positive appearance can still contain a wrong detail. Do not silently turn a passing mention into a recommendation.

Keep every attempt, including answers with no mention and failures to collect an answer. Under a stated rule, an evaluable answer with no mention belongs in the denominator. An access failure does not establish absence in an answer you never inspected. Report it separately rather than deleting the record or calling it a negative answer.

In the fictional exercise below, three of four evaluable answers contain a mention: 3 ÷ 4 = 75%. Two of those four contain an owned citation: 2 ÷ 4 = 50%. One additional attempt failed. Those are arithmetic summaries of the supplied records, not market estimates, traffic figures or evidence that a method worked.

Repeat a stated method if you want to compare collection windows. Preserve dates, prompt-set versions, product/mode changes and missingness. A stable method helps interpret differences, but does not freeze the platform or remove every outside influence. A before/after increase alone cannot isolate the effect of a rewrite.

Use the records to choose inspectable actions: verify an inaccurate business detail, improve an ambiguous source page or investigate an access problem. Keep manual answer observations separate from Google’s current Generative AI Search impression report, recorded website referrals and actual outcomes. Report what each dataset measures and what you could not observe, instead of combining them into one unsupported score.