How AI Decides Which Pages to Cite

What steps lie between a question and a cited page?

Three stages, documented at a high level by the engines that explain their own process, though none publish the exact ranking formula inside each step:

1. Retrieval. The engine pulls a set of candidate pages that look relevant. Google describes this directly for its own AI features: a technique "also known as grounding" that relies on "our core Search ranking systems to retrieve relevant, up-to-date web pages." Engines like Gemini and Claude that support live web search decide, per prompt, whether a search is worth running at all before retrieval happens.
2. Reading and reranking. The engine reviews the specific information in the retrieved pages, not just their ranking position, to decide what's useful for answering the question. A page that ranks low but answers the exact question directly can outperform a page that ranks higher but buries the answer.
3. Grounding and citation. The final answer is composed from what got used, with citations attached to the specific claims that came from a specific source.

Retrieval decides who's in the room, reading-and-reranking decides who gets listened to, and citation decides who gets credited. A page can succeed at one stage and fail at another — retrieved but not cited is a real, measured pattern (see below), not a hypothetical.

Which signals does public research link to citation?

Four independent studies converge on similar themes, though each measures something slightly different and none of them proves causation:

None of these are "add X and get cited." They're associations found in large data sets — correlational, not a controlled experiment. Both studies measure what already got cited and work backward to the pattern; neither tested whether changing a page's type or structure causes more citations.

Why are some pages ignored?

Three real patterns, from the sourced studies, cover most cases:

The page was never retrieved. If a page can't be crawled, isn't indexed, or doesn't surface for anything close to the actual question, it's never a candidate — nothing about reranking or citation matters yet. This is a technical-access problem, not a content-quality one.

The page was retrieved but outranked for relevance. The content exists and is crawlable, but something else answered the specific question more directly, more currently, or in a format the engine favored for that intent (a listicle for a commercial "best of" question, a product page for a transactional one).

The page was retrieved constantly and still barely cited. This is the pattern DeltaV's study surfaced most clearly: across three major health publishers, pages were retrieved frequently but cited at rates of 0.55, 0.33 and 0.24 citations per retrieval, while a more specialized competitor (Mayo Clinic) reached 1.79. The study calls this a "trust gap" — breadth and general visibility don't guarantee the engine treats a page as the authoritative answer once it's reading the content. Comparison pages showed the reverse problem: only 4.1% of total citations despite the highest citation rate once retrieved (1.87, about 45% above average) — a format that works well when used, but isn't produced or surfaced as often as it seems to deserve.

Which signals can a team influence?

Realistically, three of the four:

• Page type and format, matched to how buyers at that funnel stage actually ask. A listicle-style "best of" page for commercial-intent questions; a direct product page for transactional ones; a thorough article for informational ones. See which content types AI engines cite most for the fuller breakdown by industry and intent.
• Freshness, in the literal sense of keeping facts current and visibly dated — pricing, availability, anything with an "as of" quality to it.
• Specificity and extractable facts, the same discipline covered in which AI engines to prioritise: state the number, the constraint, the direct answer, rather than making a reader (or a model) infer it.

Authority, as DeltaV's study frames it, is harder to influence directly — it's closer to an earned reputation signal than a markup change. What a team can do is make sure the content that does exist is specific enough to be the clearly-better answer to the exact question, rather than a broad page that's merely relevant.

How to diagnose a page that is never cited?

A quick diagnostic, symptom to likely cause to what to check:

A page that's never cited for you but regularly cited for a named rival on the same questions is a distinct diagnostic worth running on its own — see citation gap analysis for how to compare the two side by side.

The image below shows what a citation breakdown looks like once you can see it directly — which domains got cited, how often, and at what position, split by source type.

FAQ

Is there a confirmed citation algorithm any AI engine follows? No engine publishes one. What's documented is a retrieve-read-cite pipeline, with the exact ranking inside it undisclosed. Treat any claimed "formula" from a third party as a correlational finding, not the real mechanism.

Does ranking well on Google guarantee an AI citation? No. About 80% of AI-cited URLs don't rank in Google's top 100 results at all for the same query, per Ahrefs' 15,000-query study — retrieval for AI answers and traditional search ranking are measurably different processes.

Does adding more structured data (schema) help get a page cited? Google states directly that structured data "isn't required for generative AI search, and there's no special schema.org markup you need to add" for its own features. That's specific to Google; it doesn't rule out a benefit for other engines, but no sourced study in this post found one.

Why would a page get retrieved constantly but cited rarely? DeltaV's 25,337-citation study found exactly this pattern for broad, high-traffic health publishers versus a narrower, more specialized competitor — described as a "trust gap." Being found isn't the same as being treated as the best answer.

What's the single most useful page-level change for citation? No study in this post isolates one single lever with a proven causal effect. The most consistent pattern across them is matching page format to question intent (listicle for commercial, product page for transactional, article for informational) combined with a clearly stated, current, specific answer.

What to do next

1. Confirm your most important pages are actually retrievable — check crawler access before assuming it's a content problem.
2. Pick your three highest-intent pages and check whether their format matches how buyers actually ask (listicle, article or product page) for that intent.
3. Audit for staleness: any page with a price, a feature list or an availability claim older than a quarter should show a visible, current update date.

ContextOwl sells one-time AI visibility reports. The AI Citation Snapshot is a free, single-question version of the same check.