AI citation tracking means recording when an answer visibly links to a source, together with enough context to repeat the observation. Define what counts before collecting results: a linked source URL, an unlinked brand mention and a visitor session are different events. The method below is a proposed editorial measurement workflow, not a proprietary visibility score.
Define the observation precisely
Choose one answer product and one research question. Record the product name, displayed model or mode if available, language, region when known, date, prompt and whether the conversation was fresh. If a setting is unavailable, mark it unknown rather than assuming it matches your previous run.
For each answer, distinguish whether the brand is named, whether the answer links to the site, which page is linked, and whether that page supports the associated statement. A homepage link next to a pricing claim is not necessarily evidence that the claim is supported. Visit the destination and check the relevant passage.
Google’s generative Search guidance describes eligibility and its limitations. It does not make a sampled answer an assurance of future placement. Keep your observations scoped to the product, prompt and time actually tested.
Build a small, stable question set
Select questions representing real reader tasks rather than wording designed to make your brand appear. Include informational questions, comparisons and practical troubleshooting if those match your audience. Record why each question belongs in the set, and keep the same wording during a measurement window.
Do not quietly replace questions with ones that produce more favorable results. If your audience changes, start a new version of the question set and retain the earlier version. Similarly, record failed runs and unavailable answers instead of dropping them from the denominator.
For each run, save a limited evidence record: the answer context needed to interpret the citation, the source URL and your support assessment. Remove personal prompts or private business information from material intended for a public report.
Example: report a sample without inventing market share
Suppose a fictional documentation team tests ten preselected questions on three dates. Its site appears as a linked source in six of the thirty successful answers. The team can report six observed linked-source appearances in that sample, with the question set and dates attached.
It should not call the result twenty percent of all AI searches, a market-wide ranking, or six customer visits. The sample was deliberately selected, repeated answers may be related, and users were not observed clicking the links. If two answers point to an outdated documentation version, that is a concrete maintenance finding even if the overall appearance count looks encouraging.
Separate change detection from causation
Use repeated observations to identify questions worth investigating. When a citation disappears, first check whether the answer still addresses the same task, whether the source URL moved, and whether the content changed. A changed answer interface or product mode can also make two observations less comparable.
An editorial revision followed by more citations is a sequence of events, not proof that the revision caused the change. Use the content experiment guide to plan comparisons before editing, and measure referral traffic separately when analytics evidence is available.
Avoid misleading presentation
Do not combine linked citations with unlinked mentions in one unlabeled total. Do not invent search volume weights for the prompt set, and do not present a manually selected screenshot as representative performance. Publish the sampling method alongside the result so another reader can understand its limits.
Can a citation exist without a recorded visit?
Yes. A visible source link is an answer observation. A visit requires separate traffic evidence, and analytics may not capture every referral.
Does AI Crawl Score monitor citations?
No. Its current limited public-page checks provide technical observations. This tracking workflow requires your own records or a separately evaluated measurement service.