How to measure AI visibility without false precision

Start with an operating dictionary. Accessibility means a successful response and meaningful HTML. Fetch is an attributable crawler request. Mention is the brand or entity appearing in an answer. Citation is a displayed link. A relevant citation supports the nearby claim. Referral is a visit from the answer. These events can differ in time and do not mean the same thing.

Section sources:[1] Google Search Central

Cover intents such as definition, comparison, selection, troubleshooting, and freshness. Store wording, expected entity, relevant URLs, and scoring rule for each. Do not rewrite wording after a poor answer, or you compare different experiments. Keep separate panels for languages, countries, and web-search modes; a translation is not an equivalent prompt without checking.

Section sources:[1] Google Search Central

Before a run, record surface, model or product label, locale, country, mode, timestamp, relevant account state, prompt version, and repeat count. Save the full answer, URLs, and a report snapshot. For server logs retain timestamp, URL, status, user-agent, and request ID, but never publish tokens or personal data. A consistent protocol matters more than many incomparable observations.

Section sources:[1] Google Search Central

For every answer mark mention yes/no, citation yes/no, relevance yes/no/unclear, tracked URL exact/other/none, and factual error presence. Two reviewers should independently code a sample; resolve disagreements using a prewritten rule. Do not count a platform domain as a tracked-page link without the exact URL. No link does not mean no knowledge, and a link does not mean endorsement.

Section sources:[1] Google Search Central

Show numerator, denominator, and period: “a relevant citation in 8 of 40 checks” only when actually counted. Separate unique prompts from repeats. Do not average languages and surfaces serving different users. For a small panel, provide raw rows and an interval only when a suitable model supports it; a descriptive percentage is not a causal effect.

Section sources:[1] Google Search Central

Between pre and post runs, the index, interface, model, prompt, seasonality, or source availability may change. Use a control URL, control prompts, and page version where possible. Compare identical protocol fields and mark missing runs. “More citations were observed after editing” is an observation; “editing caused growth” requires a stronger design and remains a hypothesis without controls.

Section sources:[1] Google Search Central

Documentation from Google, Bing, Gemini, Claude, and OpenAI describes separate surfaces and technical conditions, not one universal ranking of all answers. State coverage and never present one provider slice as the market. After measurement, choose an action: fix accessibility, clarify the entity, strengthen evidence, repeat the panel, or make no change. Tie each recommendation to an observation and confidence.

Section sources:[1] Google Search Central

Monitoring asks what was observed; diagnosis asks which change should be tested. A lower exact-link rate may coincide with an interface change, a different prompt panel, or page unavailability. Keep both layers in the ledger: raw event and working hypothesis. Give every hypothesis a test: rerun the same prompt, compare a control URL, check status, and open the source. Do not rewrite content before ruling out a surface change.

Section sources:[2] Google Search Central Blog[5] Google Gemini Help[6] Anthropic Support

A server log can show a request and status, but it does not prove that the URL was used in an answer. An answer snapshot can show a link, but not when or how the platform obtained the document. Link events only through an explicit identifier or close timing, and label reconstructed links as probable. In a public report do not expose IPs, tokens, cookies, or private parameters; publish aggregates and the documented protocol.

Section sources:[3] OpenAI Developers[4] Microsoft Learn

Each panel release should state period, coverage, run count, unique prompt count, missing runs, coding rules, and a raw-results table. After the table give three layers: observation, interpretation, and next action. An action might be checking canonical and HTML, not making prose more “AI-friendly.” If the difference is below a predeclared threshold or the panel is small, leave the conclusion uncertain. This report can be repeated and challenged rather than merely displayed as a polished chart.

Section sources:[1] Google Search Central

Check that every row has prompt version, locale, surface, timestamp, and denominator. Compare the exact URL with its canonical, and the source domain with the actual page shown to the reader. Count unknowns and missing runs separately; never silently remove them from a rate. Read several complete answers so relevance is not reduced to a name match. In the conclusion name alternative explanations and one next measurement. If the interface changed, start a new baseline instead of merging incomparable waves.

Section sources:[1] Google Search Central[5] Google Gemini Help