How to Check Whether AI Search Citations Support Your Content

An AI answer citing your page is not necessarily using it accurately. To check, take one claim at a time: identify the citation attached to it, open the destination, and find the passage that supports the claim as written. A visible citation is a pointer, not proof that its destination supports the sentence, as this review of citation studies puts it.

This is a different question from whether your publisher was mentioned or received a visit. A citation is a source reference attached to an answer, such as a linked marker or source card. A mention may name your publication without pointing to supporting material. A referral is a visit through a link. Record those observations separately: neither a mention nor a referral tells you whether an answer faithfully represents a page.

Capture the answer before checking the source

Start with reader questions that a specific published page is meant to answer. For each run, record the exact question, product, displayed model or mode, date, locale and relevant session settings. Save the complete answer—not just the sentence containing your name—along with its citation markers, destination URLs and copies of the pages you open. If the interface puts several references beside a paragraph, preserve their placement. A domain list collected afterward cannot show which source was presented for which claim.

For example, imagine an editorial team checking its fictional returns page. Its question is, “Can I return an opened item?” One captured answer says, “Customers can return any item within 30 days,” and links to the page. The saved page says unopened items can be returned within 14 days. The audit needs the answer’s exact wording, the linked URL and that qualifying passage. Calling the answer cited does not settle whether “any item” or “30 days” is supported.

This level of care matters even for a narrow source-identification task. In the Tow Center’s AI search study, tested chatbots collectively answered incorrectly on more than 60 percent of queries. The test presented news excerpts and asked tools to identify the corresponding articles; it was not a measurement of the error rate for every AI answer or publisher page. Its authors also note that each excerpt was queried once and that the same prompt could produce a different response on another run.

Give each claim its own finding

Break the captured answer into checkable claims. For each, note the attached citation, where its link actually lands, and the exact passage that supports or conflicts with the wording. Open the page rather than accepting a relevant-looking title: the question is whether it contains the specific information claimed. Check dates and qualifications when the claim is about a current policy, price or event.

A compact review sheet can use these fields:

Field	What to record
AI claim	Exact sentence or separable assertion, including qualifiers
Citation	Marker or card attached to that claim; mark uncited if none is shown
Destination	Linked URL, page title, publisher and whether the page was available
Source passage	Exact supporting or conflicting text, with its date where relevant
Finding	Supported, unsupported, distorted, attribution error, broken link or uncheckable

Keep the finding precise. Unsupported means the inspected page does not establish the claim; distorted means the answer drops or changes a material qualification, as in the fictional returns example. If the page supports the statement but the reference names the wrong publisher or points to a republished copy rather than the original, record an attribution error instead. The Tow Center observed links to syndicated and unofficial versions of articles rather than original publisher pages, so source identity deserves its own check. If a destination fails to open, record a broken link; do not infer what an unavailable page says. Mark a claim uncited when no source is attached, and uncheckable when the saved material cannot resolve it. A correct statement can still have a wrong citation.

Citation coverage is another, separate count. An answer might make five checkable claims, attach references to two, and support only one through its linked passage. Report both denominators—claims reviewed and cited claims inspected—rather than folding everything into a single “accuracy” percentage. Also keep a row for relevant publisher pages that received no mention or citation in a captured answer. That is a visibility observation, not proof that a crawler was blocked or that the page was never retrieved. The Tow Center’s publisher sample included sites that permitted search bots via robots.txt and sites that blocked them; the researchers could not observe every alternative blocking control. Investigate access as its own question rather than diagnosing it from an answer alone.

Report the mismatch you can demonstrate

A useful report names the tested products and modes, capture dates, locales, exact prompts, number of runs, claims reviewed and citations inspected. Preserve each run rather than replacing earlier answers with the latest one. If you repeat the same question, a changed result establishes what appeared in the new capture, not that every reader now sees it.

For a correction request about the fictional returns answer, include the AI claim (“any item within 30 days”), the cited URL, the passage specifying unopened items and 14 days, and the two mismatches. If the link instead led to another publisher’s copy of an otherwise accurately quoted policy, describe that attribution problem rather than alleging the policy is false. A later check should ask whether the particular mismatch remains in the newly saved answer and destination. It cannot, on its own, establish why the earlier answer appeared or whether a correction will persist.