How Do Answer Engines Choose Sources—and What Can Publishers Actually Know?

A displayed citation tells you which linked source an AI answer attributes material to. It does not show every page the system considered, where that source ranked among candidates, or why it was chosen. Publishers can learn from the answer in front of them, but a citation is an output—not a disclosure of the selection rules.

That distinction matters when someone asks, “How do I get cited?” Before looking for a tactic, separate four things: whether a page can be found for a search request, whether it was considered for this answer, whether the answer cites its URL, and whether the answer merely mentions its brand. Only the last two may be visible in the answer itself.

What a link and a brand mention establish

Imagine a hypothetical answer to “Can I machine-wash a North Quay rain jacket?” It mentions North Quay by name and links to an independent care guide. The visible result establishes that the brand was mentioned and that the answer attributes material to the guide. It does not establish that North Quay’s own care page was retrieved, rejected, inaccessible, or used without a citation. Nor does the link, by itself, prove that the guide supports every claim in the answer.

Change the example so the link points to North Quay’s care page. Now the publisher has an observed citation to its URL—not evidence that its page outranked the independent guide, or that either source will appear for another question. A brand mention and a cited publisher URL are different observations; neither is the same as a recorded referral or an attributed conversion.

If the concern is whether an answer accurately represents its linked page, that is a separate question from how the link was selected. See How to Check Whether AI Search Citations Support Your Content for that investigation.

A useful sequence, not an engine diagram

One way to reason about source selection is to picture discovery or retrieval, evaluation of candidate material, selection of relevant passages, and composition of an answer with possible citations. Citevera describes a four-stage framework using retrieval, reranking, citation eligibility, and final composition. Len van der Hof offers a related diagnostic model covering access, understanding, relevant passages, attribution, and value relative to alternatives. These are authorial frameworks for asking better questions, not verified diagrams of every product’s implementation.

The sequence helps locate the uncertainty. If a page cannot be discovered or fetched for a search request, it cannot contribute as a retrieved page in that request. If it can be fetched, that clears only an access hurdle; it does not show that the page was retrieved for a particular answer, selected as a source, or cited. Retrieval or search access is also a different question from whether a provider may use material for model training.

There are product-specific eligibility conditions, but they should not be inflated into selection formulas. Martin Zialcita’s account reports that a page must be indexed and eligible for a snippet to be considered for Google AI Overviews. That describes a prerequisite for consideration, not a way to secure a citation. The same account notes that major providers do not publicly document the specific signals used to order AI-search retrieval candidates. An accessible, eligible page can therefore still be absent from the displayed sources without the answer revealing why.

What citation patterns can—and cannot—tell us

Published observations are more useful when their sampling boundaries stay attached. In XFunnel’s study, the authors generated the questions and analyzed 40,000 answers from ChatGPT, Perplexity, and Gemini, capturing 250,000 unique outbound links. They also report that interface models occasionally changed during measurement; the measurement period is not specified. This is a substantial observed set, but not a current, universal rule for choosing sources.

In that set, earned third-party sources were the most frequent citation type across the three engines. The authors also report different category patterns across their buyer-journey question groups: earned sources were prominent in early-stage questions, user-generated content increased in comparisons, and owned or competitor domains appeared more in later-stage groups. Those distributions describe the answers to those generated questions. They do not demonstrate that earning coverage causes citations, that every comparison question needs a forum source, or that an engine applies a fixed preference for a category.

This is the useful boundary between observation and mechanism. A cluster of cited domains may suggest questions worth investigating. It cannot tell you which unseen candidates competed for each answer or the weights, if any, assigned to their features. Even a repeated citation pattern is still a pattern in outputs, not a look inside the selection decision.

The question to ask of an unexplained absence

Suppose North Quay’s care page is indexed and eligible where that condition applies, yet an answer cites only the independent guide. The supported conclusion is narrow: in the answer observed, the guide got a link and North Quay’s page did not. “The engine prefers third-party sites,” “our page lacks the right markup,” and “the guide was ranked higher” are competing explanations, not findings from that answer.

Publishers can use the conceptual sequence to decide what kind of evidence they need next. An access check addresses whether the page clears an access hurdle. A saved answer establishes what was displayed for a particular question and check. Neither reveals the complete candidate set. If the real goal is to count appearances across checks, What Can AEO and GEO Tools Actually Measure? is the more useful next question. If the goal is to explain why one source won, the honest answer may remain: the displayed citation does not say.