Being Cited by an AI Doesn't Mean Your Details Reached the Answer

When a page mentioning your business shows up in an AI assistant’s reference list, it’s easy to count that as a win. In reality, a citation only proves the system found a source link. It doesn’t guarantee the AI actually named your business in its response, got your location right, or passed along the phone number a customer needed.

Appearing in a citation pool and appearing in the final text are two entirely different things. Any serious visibility tracking needs to treat them that way.

A study examining Chinese-language generative search looked directly at this disconnect by mapping generated answers against their visible citations and the underlying crawled web pages. The researchers ran 614 controlled queries across both the web and mobile app interfaces of four different platforms, repeating each combination three times.

In their dataset, only 8.3% of the brand candidates found in the visible citation pool actually made it into the written answers. For cited sources containing contact information, the carryover rate was 12.4%. Keep in mind, these numbers aren't universal laws of AI behavior—they reflect how this specific study defined candidates and matched text. Still, they clearly show that getting cited is not the same as getting featured.

The Reference List Isn't the Whole Story

A major limitation in tracking AI search is that researchers can only see the citations the interface chooses to display. That visible list isn't a complete log of every document the retrieval system pulled behind the scenes, nor does it represent the model's pre-existing training data.

This helps explain why source pages and final answers often don't line up. An AI might cite a webpage to back up a single sentence while ignoring everything else on the page. If your brand is mentioned on that page, it might just be irrelevant to the user's prompt. Similarly, the model has no reason to copy over a street address just because it happens to be readable in the source text.

The researchers also saw the opposite problem: answers sometimes included brand names and contact details that couldn't be found in the visible citations or the crawled page text at all. When this happens, we shouldn't jump to conclusions about a single hidden mechanism. It simply shows that the displayed links, the text a third-party crawler manages to extract, and the data the AI actually uses aren't always identical.

Because of this, counting citations isn't enough. A proper audit has to look at both sides: what the answer actually claims, and what the linked page visibly supports.

How Site Quality Fits Into the Picture

The researchers also tested which factors best predicted whether information would move from a cited page into the final answer. Prominent predictors included how well the content fit the query, how often the information repeated across multiple sources, and the specific semantic role that information played. A composite site-quality score the researchers tested wasn't the top factor.

However, it would be a mistake to read this as "site quality doesn't matter for AI visibility." In fact, the paper notes that their composite score had a moderately positive link to brand selection. More importantly, their findings only apply to the specific quality metric they built, not the concept of quality as a whole.

Their metric bundled together rough proxies like indexed page counts, third-party domain weights, and registration data. Those high-level site signals can't tell you whether a specific paragraph actually answers a user's question, or whether a specific detail belongs in the final output.

It's also crucial to separate observational predictions from cause-and-effect. Just because a metric helps predict which details get selected doesn't mean tweaking that metric on your site will force the AI to include you. The study’s authors rightly treat their practical takeaways as starting points for controlled experiments, not guaranteed formulas.

Don't Mix Web and App Data

By testing four platforms across both web browsers and mobile apps, the study effectively tracked eight distinct environments. That’s a smart approach to copy, because a platform name alone may hide differences between its web and app interfaces. A platform's website and its mobile app might retrieve, filter, or display information differently.

If you lump web and app results together in your reporting, a simple shift in how you collect data can look like a sudden spike or drop in brand visibility. Always log which interface generated which answer, and compare apples to apples. The same rule applies when a single product offers multiple search modes (like a fast chat versus a deep research toggle).

Content age requires the same kind of context. While the study mapped out the age distributions of cited pages, those curves don't mean every webpage has a strict expiration date. An older article might be useless for a query about current pricing, but completely fine for explaining background history.

How to Audit a Specific Fact

Instead of obsessing over citation counts, try tracing a single piece of information a customer actually needs—like your brand name, service area, physical address, or a specific product requirement.

Start by asking the AI a question where that specific detail naturally belongs in a helpful response. Save both the full text of the answer and the links it provides.

Next, open the cited page and find the exact passage that backs up the claim. Is your fact actually on the page? Is it clearly labeled, and did the AI copy it over without twisting the meaning?

If your page was cited as a source, yet your fact was left out of the answer, ask yourself if the prompt genuinely required that detail. On the flip side, if the AI includes your fact without any backing from the cited pages, log it as an unresolved support gap rather than guessing how the model knew it.

If the source text on your site is messy or vague, testing clearer phrasing and better formatting makes sense. Just keep your eyes on the real goal: getting the right fact accurately placed into the written answer. Racking up more footnote links won't tell you if you've fixed the problem.

Note: These takeaways are based on an observational study of Chinese-language web and app interfaces using a specific text-matching method. While the study's site-quality score showed a positive link to selection, these findings don't prove how page edits will behave across other AI systems.