Pages found vs. pages cited: why most sources an AI reads never appear in its answer

A page is found when an AI assistant's search surfaces it while working on an answer. A page is cited when the finished answer points to it as a source. The two are very different counts: an assistant can look at a long list of pages and cite only a handful, so most of what it reads never appears in what you see.

Why are there two numbers at all?

Because answering is a funnel. The assistant runs searches, which return a pool of results. It reads or skims part of that pool. It writes an answer from some of what it read. It then attaches citations to some of what it used.

At each stage, material drops out.

• Found: returned by the search.
• Read: opened or taken into the assistant's working context.
• Used: shaped a sentence in the answer.
• Cited: linked or listed as a source.

From the outside, you can often see only the last of these. Some providers also report the first. The middle two are generally invisible.

How big is the gap?

It depends on the assistant and the question, and it can be large. Citewake's Edition 0 results measure it, and the figures are in the findings guide and the editions. The pattern to expect without the numbers: an assistant that reports a broad set of pages found tends to cite far fewer, while an assistant that cites generously may report little beyond what it cites.

That is why you should never compare "sources" across assistants without asking what is being counted.

If a page is found but not cited, did it do anything?

Possibly, and you usually cannot know. A found page might have:

• shaped the answer without being credited;
• confirmed what the assistant already believed;
• been skimmed and dropped because another page said the same thing more clearly;
• been irrelevant.

This matters for anyone trying to explain why a brand was or was not recommended. The honest statement is narrow: "this page was cited" is a fact about the answer, while "this page influenced the answer" is an inference. Citewake's method treats a brand found on a cited page as exactly that and no more: the brand was in its fetched copy of a page the answer cited. It does not claim the assistant read that passage. The methodology states this limit.

Why does cited matter more for visibility?

Because the citation is what a person can click, and because it is the assistant's public statement of where its answer came from. Being cited can bring a visit, and it signals that the assistant treated your page as worth pointing to. For the commercial side, see what an AI citation is worth to a business.

But "cited" can be a weak target on its own. Citations often go to explanatory pages, documentation and roundups rather than to the pages of the brands being recommended. So a brand can be recommended without being cited at all, and a site can be cited often without being recommended. We cover the difference in mentioned, recommended or cited.

What does this mean for how you measure?

Three rules help.

1. Count each thing separately. Report pages found, pages cited, brands named and brands recommended as different measures. Never merge them.
2. Say what the provider reports. If an assistant does not report pages found, write "not reported", not zero.
3. Keep the denominator fixed. A rate is answers with a citation divided by all valid answers, not a share of whatever happened to be reported.

These rules are the reason a lab bothers to keep each assistant's evidence apart rather than blending it into one score. Providers report different things, and pretending otherwise would make the comparison look cleaner than it is.

What can you do about it as a site owner?

You cannot make an assistant cite you. You can make your page easier to use and easier to credit.

• Answer a specific question on the page, early. Assistants quote what is easy to lift.
• Put facts where they can be checked. Plain statements with clear sources beat vague claims.
• Be on the pages that get cited. Often those are third-party pages. See the kinds of pages AI assistants cite most.
• Let the crawlers in. See AI crawlers and robots.txt.

Why do some assistants cite so much more than others?

Citation habits are part of each assistant's design. Some attach a source to nearly every claim. Others cite a few pages at the end, or only when the user asks. Some show everything they retrieved and others show only what they used. So a high citation count is not proof of a more thorough answer, and a low one is not proof of a thin one. When you compare assistants, compare like with like: for each, say what the number counts, and avoid ranking them on a count that means something different in each. This is also why a lab keeps each assistant's evidence in its own column, rather than adding them into one tidy total that nobody could interpret.

How can you see the gap yourself?

Use an assistant that shows its sources. Ask a buying question and count the sources listed. Then, if the assistant has a view of its searches or a "sources considered" panel, count those. Compare the two numbers across a few questions. The gap you find is the practical reason to ask what a "source" count actually counts. To get a first read on how assistants see your own site, you can try the Citewake checker.