How do AI answer engines decide which sources to cite?

An AI answer engine does not choose a website. It chooses a passage. The question is converted into a vector, an index returns the chunks of text whose vectors sit closest to it, the highest-scoring chunks are assembled into a working context, and a language model writes prose from that material and attaches citations to the passages it leaned on. Every practical decision about how to structure a page follows from that one fact: the unit of selection is a few hundred words, not a document.

This piece is the full mechanism, with the research behind each step named. It is longer than anything else on this site because it is the thing everything else depends on.

What happens between a buyer's question and a citation?

Four steps, in order. Naming them separately matters, because a company can be invisible at any one of them for reasons that look identical from outside — you simply are not named, and no report tells you which step failed.

Step one: the question is expanded, not just matched

A buyer types one question. The system does not necessarily search for that one string. It commonly rewrites the query into several related queries, retrieves for each, and pools the results — the practice Mike King's AI Search Manual discusses at length under query fan-out. The practical consequence is that you are not competing for a phrase. You are competing across a cloud of related phrasings, several of which the buyer never typed and you will never see.

This is why a keyword list is a poor planning tool here and a list of buyer questions is a good one. The questions survive rewriting; the exact keywords do not.

Step two: passages are retrieved, not pages

Retrieval-augmented generation — the architecture underneath all of this — was set out by Patrick Lewis and colleagues in the 2020 NeurIPS paper that named it, combining a language model's internal knowledge with a searchable external index the model consults at answer time. Documents in that index are split into chunks and each chunk is embedded independently as a vector.

So the thing being scored is a fragment of your page, evaluated without the paragraph above it, the heading two sections up, or the sentence that defined your acronym. If a section only makes sense in place, it does not make sense at the moment it matters.

I have deliberately not quoted a chunk size here. Numbers between roughly 200 and 800 words circulate freely, but no major engine publishes its chunking parameters, and repeating one as fact would be inventing precision. What is safe to say is directional: sections should be short enough to survive being lifted whole, and self-contained enough to be worth lifting.

Step three: the context window is assembled

The top-scoring passages, drawn from several different sites, are placed in front of the model together. Your passage is now sitting alongside competitors' passages answering the same question, and the model is choosing between them on the evidence in front of it.

This is the step people underestimate. Being retrieved is not being cited. Retrieval gets you into the room; what happens next depends on whether your passage answers the question more directly, more specifically and more verifiably than the three others beside it.

Step four: the answer is generated and citations are attached

The model writes prose and attributes claims back to the passages it drew on. Attribution is the reason entity resolution matters: a citation is an act of naming a source, and a system that cannot establish which organisation a page belongs to has a weaker basis for naming it. That is the argument for structured data and a corroborated identity, covered in how a model decides your company exists.

Why can a page rank in the top ten and never get cited?

Because ranking is a page-level judgement and retrieval is a passage-level one, and the two have measurably come apart. Ahrefs studied Google's AI Overviews across 863,000 keywords and 4 million cited URLs and reported in March 2026 that roughly 38% of cited pages also appeared in the top ten organic results for the same query — against about 76% when they ran the same study in July 2025.

Read that carefully in both directions, because most write-ups only read it one way. The overlap halved, which means ranking no longer predicts citation the way it did. It did not vanish, which means roughly two in five cited pages were still strong organic performers. Both facts are load-bearing: abandoning search fundamentals removes a floor you are standing on, and relying on them alone is betting on a relationship that weakened by half in eight months. The fuller treatment is in what the Ahrefs study actually measured.

Does any of this happen if a crawler cannot read the page?

No, and this is the failure that makes every other consideration irrelevant. Vercel, working with research partner MERJ, analysed AI crawler behaviour in December 2024 and found the major AI crawlers do not execute JavaScript: GPTBot requested JavaScript files 11.5% of the time and ClaudeBot 23.8%, and neither ran them. Gemini and Applebot are the exceptions, because both reuse rendering infrastructure built for classic search.

If your pages assemble themselves in the visitor's browser, most of these systems receive a near-empty document. No amount of passage craft fixes that. The one-command test is in how to check what an AI crawler actually sees.

What does the research say actually influences whether you are cited?

There is one controlled study worth taking seriously, and it is worth being precise about what it did and did not show.

Pranjal Aggarwal and colleagues at IIT Delhi and Princeton University published "GEO: Generative Engine Optimization" at ACM SIGKDD 2024. They tested nine optimisation methods across GEO-Bench, a benchmark of 10,000 queries. The three best performers were adding quotations, adding statistics and citing sources, reported as a 30–40% relative improvement on the paper's Position-Adjusted Word Count metric and 15–30% on its Subjective Impression metric. The authors also found effectiveness varied by domain, which argues against treating any of it as a universal recipe.

Three caveats that rarely survive the journey into agency marketing. The metric measures prominence inside a generated answer within a research benchmark — not traffic, not revenue, not a result at a company. The gains were relative, not absolute. And the effect depended on the quotations and statistics being real: engines cross-reference, so a figure with no traceable source weakens the passage it was meant to strengthen.

Where do the sources disagree, and who is right?

This is the part most explainers skip, and it is the most useful part.

The GEO paper demonstrates that specific writing changes measurably shift how prominently a source appears in a generated answer. Google's own documentation points in what looks like the opposite direction: its guidance on AI features in Search states there is no special markup or additional optimisation required for AI Overviews and AI Mode, and that the ordinary rules of making content discoverable apply.

Both are true, and reconciling them tells you what to spend money on:

• Google is describing the eligibility layer. There is no AI-Overviews schema type, no magic tag, and nobody can buy their way in. Any vendor selling "AI Overview markup" is selling something that does not exist.
• The GEO paper is describing the selection layer. Given a pool of eligible passages, the ones that are specific, attributed and quotable get drawn on more often. That is not a trick — it is the same quality signal a careful human reader applies.

So the honest position is narrower than the marketing and broader than the scepticism: there is no separate technical channel to optimise, and there is a real, measured difference between a passage that answers with evidence and one that asserts.

What is the evidence-graded version?

Sorted by how much weight the evidence actually bears, rather than by how often the tactic is sold:

Which sources do engines actually draw on?

Yext analysed 6.8 million citations across 1.6 million AI answers from Gemini, OpenAI and Perplexity and reported in October 2025 that "86% of citations come from sources brands already control" — 44% first-party websites, 42% listings and profiles, with the remainder from reviews, social, news and forums.

One analytical note the headline hides: this is descriptive, not causal. It tells you where citations came from across a large sample. It does not tell you that adding another directory listing produces a citation, and reading it that way leads directly to mass submission, which manufactures the inconsistent descriptions that make an entity harder to resolve. The ordered version is in which sources to fix first.

How do I apply this to a page I am about to write?

1. Confirm the page exists without JavaScript. Everything below this line is wasted until it does.
2. Write the heading as the question a buyer would type. The heading travels with the chunk and is often the strongest signal in it.
3. Answer in the first sentence beneath it. A retrieved passage that opens with preamble has spent its most valuable position on nothing.
4. Resolve every pronoun inside the section. "This approach also helps" is unusable when the referent did not travel with the chunk.
5. Name the source in the same sentence as the figure. Not a footnote, not a link list at the bottom — the same sentence, because the sentence is what gets retrieved.
6. Keep sections short enough to survive being lifted whole. Roughly 40 to 150 words is a working range, not a rule handed down from an engine.

The worked version of this, with before-and-after examples, is in how to write a page an AI engine can actually quote.

What cannot be controlled?

Source selection at the moment of generation. Answer engines are non-deterministic: the same question can return different answers to two people on the same day, and operators change retrieval behaviour without notice or announcement. What is controllable is whether the business is identifiable, whether its content exists in retrievable form, and whether the third-party sources these systems lean on describe it accurately.

Anyone offering a guaranteed position in an AI answer is describing something they do not control. That is the single most reliable disqualifying question you can ask, and there are eight more in what to ask before hiring anyone for this.

What to do about your own situation

Before commissioning any of this, establish empirically whether the engines your buyers use currently name you, which sources they cite when they name someone else, and whether your pages are readable to a crawler that does not run JavaScript. That is a measurement, not a strategy, and it costs nothing but the asking.

The service it belongs to is getting found when buyers ask AI.

Run a free AI visibility check