When AI Cites a Page, Which Details Reach the Answer?

Your business appears on a page cited by an AI answer. That is evidence of a source connection. It does not tell you whether the answer named your business, included the right location or carried over the contact detail the customer needed.

Those are separate observations, and a visibility report should keep them separate.

A study of Chinese-language generative search linked answers, their citations and crawled source pages to inspect that gap. It covered 614 controlled queries across the web and app interfaces of four platforms, with three replications for each query–platform–interface combination.

The researchers reported that 8.3% of brand candidates in the observed citation pool were selected into answers. For sources containing contact information, the reported contribution rate was 12.4%. These figures concern the study's defined candidates and linkage methods. They are not universal probabilities that any brand or phone number on your website will be used.

A reference list shows only part of the path

The study's citation pool consists of references visible to the researchers. It is not a complete record of every document the system retrieved internally or everything available to the model.

That matters when interpreting a source-to-answer gap. A page can appear as a reference while only one of its claims is used. A brand on that page may be incidental to the question. The assistant may have no reason to include its address, even if the address is readable.

The reverse mismatch also occurred. Some brand and contact-information exposures could not be matched to the contemporaneous citation pool or crawled body text. That does not establish a single hidden mechanism. The exposed references, the crawler's extraction and the information the system actually used need not be identical.

A useful audit therefore records both sides: what the answer said and what the visible source supports. A citation count cannot stand in for that comparison.

One quality score had a positive association

The paper tested several predictors of what surfaced. Content fit, recurrence across sources and the information's semantic role were prominent in the predictive models. A particular composite site-quality score was not the leading predictor.

It would be wrong to shorten that result to “quality does not predict visibility.” The paper describes the composite as a moderately positive predictor of brand selection. Its result concerns that particular score and the measured outcomes, not every possible definition of source quality.

The score combined signals such as indexed-page count, a third-party weight measure and site-registration information. Those are proxies. They cannot fully describe whether a specific page answers the question or whether a particular detail belongs in the response.

Predictive importance is also different from an editing experiment. A variable that helps predict selection is not automatically a lever you can change to produce a desired result. The authors frame practical implications as hypotheses for controlled testing, which is the right level of confidence.

Keep web and app results separate

The study examined eight platform-interface combinations. That design is useful because a platform name alone may hide differences in how its website and app obtain or present information.

If your report pools both interfaces, a change in the mix of collected answers can look like a change in your visibility. Record which interface produced each answer and compare like with like. The same principle applies when a product offers different search modes.

Recency needs similar care. The paper fitted age patterns among dated cited pages. Such a distribution does not give an individual page an expiry date. A source that is old for a question about today's price may still be appropriate for a historical explanation.

Trace one important fact

Choose a fact a customer needs: your brand name, a service area, an address or a product requirement. Ask a question where that fact genuinely belongs in the answer, then save the complete answer and visible sources.

Open the cited page and locate the supporting text. Check whether the fact is present, whether its labels remain attached and whether the answer reproduces it accurately. If the page is cited but the fact is absent, first consider whether the question called for it. If the answer includes the fact without visible support, record that as an unresolved support gap rather than inventing a retrieval story.

Clearer wording or structure may be worth testing when the source itself is ambiguous. The observable target is accurate transfer of the needed fact into the answer. More citations alone would not show that the problem had been solved.