AI Citations: What to Check and How to Measure Them

A mention, a citation and a visit are different

A brand mention names a business in an answer. A citation links to a source used or presented with the answer. A visit happens when someone reaches the website. Record them separately: a business may be named without a link, and a cited page may discuss a business other than its owner.

For measurement, define an owned citation as a link to a URL on the domain you are tracking. Preserve the full cited URL in the observation record, then normalize it for analysis without losing the original evidence. A third-party review that mentions your company is a different source category.

Check whether the source supports the statement

Open the cited page and find the passage relevant to the answer. Check the operator, service, location, price and date against an approved source. A link next to a sentence is a lead for verification, not proof that the whole answer is accurate.

These are hypothetical checks, not measured outcomes. Use the accuracy log when a statement is outdated or cannot be verified.

Make your page useful as a source

Put the answer where a reader would expect it: service scope on a service page, plan details on Pricing and company identity on About. Explain terms that are easy to confuse. Identify whether a number is a price, an example, a vendor estimate or an observed result.

Keep sources and dates close to claims that depend on them. A comparison should identify the documentation reviewed and any commercial relationship. A research observation should expose its sample and limitations. Avoid turning a small sample into a claim about every AI user.

Google's content guidance emphasizes useful, reliable information and clear authorship. Adding a link or an author name cannot replace evidence for a claim.

Access is necessary to check, but it is not inclusion

Check robots rules, response codes, canonical URLs and rendered content before diagnosing an absence of citations. A security challenge or inaccessible page can prevent a particular retrieval, but one failed request should not be generalized to all crawlers.

OpenAI documents distinct crawler purposes, including search and model-training crawlers. Use the documented controls for the intended purpose. Allowing a crawler does not guarantee a specific answer or citation.

If a source page is accessible, investigate the next questions: is it indexed where required, does it answer the prompt, is the information current, and what sources did the observed answer actually choose?

Measure a defined sample

For a fixed question set, report the number of successful observations containing an owned citation divided by the number of successful observations. Keep failed runs separate. Split branded and unbranded questions, and do not mix different platforms or execution conditions into a trend without explaining the change.

Keep a record of which pages were cited. That reveals whether a specific service guide is useful in observed answers, even when a site-wide percentage hides the detail. Pair this with search performance and actual enquiries when available; neither one citation nor one monitoring score establishes business impact.

Our measurement guide includes a dated FoundUp API record. Use the monitoring scope to specify the evidence you want to receive.