How to measure AI visibility without false precision

Start with simple definitions. A page is accessible when it loads successfully and its main text is present in the page code. A bot visit is a server request that can be linked to a known operator. A mention means the answer names the brand or subject. A source link is a link shown to the reader; it is relevant only when it supports the nearby claim. A visit means someone followed that link. These events can happen at different times and must be measured separately.

Section sources:[1] Google Search Central

Prepare prompts for different user needs: a definition, comparison, choice, troubleshooting step, or current information. For each prompt, save the exact wording, the subject you expect the answer to cover, the pages that would be useful sources, and the rule used to assess the result. Do not rewrite a prompt after a poor answer, because that creates a different test. Test languages, countries, and internet-search modes separately; check a translation before treating it as equivalent.

Section sources:[1] Google Search Central

Before each run, record the service and mode, the AI version, language, country, time, and number of repeats. Note whether the result depends on being signed in. Save the exact prompt, the complete answer, every page address shown, and a screenshot. Server records need only the request time, page address, response code, software name, and request number. Never publish access tokens or personal data. A consistent test is more valuable than many runs that cannot be compared.

Section sources:[1] Google Search Central

For each answer, record five separate facts: whether it names the brand; whether it shows a source link; whether that link supports the nearby claim; whether it leads to the target page, another page, or nowhere; and whether the answer contains a factual error. Two people should review the same sample independently and resolve differences with a rule written in advance. A link to a service homepage is not a link to the material being studied. No link does not mean no knowledge, and a link does not mean endorsement.

Section sources:[1] Google Search Central

Put the count and period beside every percentage: for example, ‘a useful source link appeared in 8 of 40 checks during one week.’ Count distinct prompts separately from repeats. Do not merge different languages or service modes when they serve different audiences. For a small sample, show the underlying rows; add a statistical range only when the method truly supports it. A percentage alone does not prove that one change caused another.

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 source links 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 services and technical conditions, not one universal ranking of all answers. Name exactly what was tested and never present one provider's result as the whole market. Then choose a concrete action: make the page accessible, describe its subject more clearly, add stronger sources, repeat the same fixed test, or make no change. Explain in ordinary words which observation supports each recommendation and how certain it is.

Section sources:[1] Google Search Central

Monitoring records what happened; diagnosis identifies which possible cause to test. A smaller share of links to the exact page may coincide with an interface change, a different fixed prompt set, or a page becoming unavailable. Keep two separate records: the observation and the working explanation. Test each explanation by repeating the same prompt, comparing a control page, checking the server response, and opening the source. Do not rewrite the content until you have ruled out a change in the service itself.

Section sources:[2] Google Search Central Blog[6] Google Gemini Help[7] Anthropic Support

A server log can show a page request and response code, but it does not prove that the page was used in an AI answer. An answer snapshot can show a link, but not when or how the service obtained the document. Connect events only through an explicit request identifier or close timing, and label a reconstructed connection as probable. In a public report, do not expose IP addresses, access tokens, cookies, or private parameters; publish aggregate results and a clear description of the method.

Section sources:[4] OpenAI Developers[5] Microsoft Learn

Each report should state its period, coverage, run count, unique prompt count, missing runs, review rules, and a table of underlying results. Follow the table with three short parts: observation, explanation, and next action. A useful action may be checking the page's primary address and source code, rather than vaguely making the prose more ‘AI-friendly.’ If the difference is below a predefined threshold or the sample is small, leave the conclusion uncertain. This kind of report can be repeated and challenged instead of merely being displayed as a polished chart.

Section sources:[1] Google Search Central

Every row should contain the exact prompt and its version, language, service mode, run time, and the total number of checks used to calculate the rate. Compare the displayed address with the page's primary address and open the source exactly as a reader would. Count unknown results and missing runs separately; never remove them silently from a percentage. Read several complete answers so a useful source is judged by meaning rather than a matching name. State other possible explanations and the next measurement in the conclusion. If the service interface changes, start a new observation series instead of merging periods that cannot be compared.

Section sources:[1] Google Search Central[6] Google Gemini Help

Google Search Console provides a filter that separates branded and non-branded queries in Google web-search reporting. Google warns that the classification can be wrong and does not affect ranking. This is a fact about a Google report, not a rule for ChatGPT, Perplexity, Gemini, Claude, or Alice.

For AI-visibility measurement, the same boundary is useful as an editorial rule. A prompt such as “what does brand X offer?” tests an answer about a named subject already named. A task prompt without the name tests whether the brand appears without that cue. Combining both groups into one rate hides two different test conditions.

Section sources:[3] Google Search Central Blog[1] Google Search Central

For a scenario prompt, record four fields: the person's task, context, constraints, and decision criterion. For example: “I want to manage a budget on my phone. I need synchronization, export, and little manual setup. Which apps should I compare before installing one?” This illustrates the form; it is not a research result.

Check whether the constraints repeat the target brand's slogan or a unique feature. If another reviewer can infer the required winner from the wording, the scenario probably contains a hidden cue. This is a test-design rule, not a ranking fact. For each group, report unique prompts, repeats, missing runs, and the underlying count separately.

Section sources:[3] Google Search Central Blog[1] Google Search Central