How do AI answer engines decide what to cite?

Most run a search, fetch a handful of top results, split those pages into passages, pick the passages that best match the question, and attribute the sentences they used back to the page the passage came from.

Why does my page rank but never get cited?

Usually because the answer is spread across several paragraphs rather than stated in one. Retrieval works on passages, so an answer that only makes sense after reading the whole page rarely survives the chunking step.

Do I need to be on page one to be cited?

It helps a great deal, because most answer engines only fetch a small number of results. Being in the top handful for the exact phrasing of a question matters more than being first for a broad head term.

Answer engine optimization

How ChatGPT, Claude, and Perplexity pick the sources they cite

A practical look at retrieval, chunking, and citation, and what each step implies for how you should structure a page you want quoted.

September 2, 2026 · 9 min read

If you want to be cited by an AI answer engine, it helps to know roughly what happens between someone typing a question and a citation appearing. The pipeline is not mysterious, and each stage has a clear implication for how you should structure a page.

What follows is the general shape of how these systems work. The specifics vary between products and change often, so treat this as mechanics rather than as a description of any one system's internals.

Stage one: the question becomes queries

The user asks something conversational. The system turns that into one or more search queries, which usually look nothing like what the user typed.

Someone asking "what should I use instead of Notion for a small engineering team" might generate queries for Notion alternatives, project management tools for engineering teams, and comparisons of specific competitors.

What this implies: the phrasings you should target are the generated queries, not the conversational question. Those look like conventional search queries, which is one more reason ordinary keyword research still applies.

Stage two: retrieval

The system runs those queries and takes results. Not fifty results. Usually somewhere between three and ten per query, because fetching and processing pages is slow and expensive.

What this implies: the top few positions matter enormously, and the drop-off below them is much sharper than in conventional search. Ranking eighth for a query might still get you occasional clicks in a search results page. It very often means not being fetched at all here.

It also implies something less obvious. Because these systems fetch a small number of pages per query, pages that aggregate many options are disproportionately valuable to them. One directory or comparison page covering twenty tools answers more of the question than one vendor's homepage. This is why listing and comparison pages show up in AI citations far more often than their search traffic would suggest.

Stage three: chunking

Fetched pages get split into passages. The exact method varies, but the effect is consistent: your page stops being a page and becomes a set of chunks of a few hundred words each.

What this implies: this is the stage that kills most product pages. If your answer to "how much does it cost" is spread across a pricing section, a FAQ, and a footnote, no single chunk contains the answer, and no chunk gets selected.

It also means your page structure determines your chunk boundaries. Clear headings with self-contained sections beneath them produce clean chunks. Long undifferentiated prose produces chunks that start and end mid-thought.

Stage four: selection

The chunks get scored against the question, usually by semantic similarity, and the best few get passed to the model that writes the answer.

What this implies: the chunk needs to be recognizably about the question, on its own, with no surrounding context. A chunk that says "it also supports this out of the box" scores badly because there's no way to tell what "it" is or what "this" refers to.

Naming the subject explicitly in most paragraphs feels repetitive when you read the page top to bottom. It's the single highest-leverage habit for being quoted.

Stage five: generation and attribution

The model writes an answer from the selected chunks and attributes sentences back to the pages the chunks came from.

What this implies: the attribution is to the source of the passage that got used. So being cited is not about being the best page overall. It's about having written the specific passage that best answered one specific question.

A modest page with three genuinely clear answers can get cited more often than a comprehensive page where everything is hedged.

What this adds up to

Five practical conclusions, in rough order of impact:

Be in the top few results for the queries that get generated. This is conventional SEO, and it's the gate everything else sits behind.
Put the answer directly under the question. Heading, then answer, in the first sentence or two.
Name your subject explicitly. Assume every paragraph will be read in isolation, because it will be.
State facts plainly. Numbers, specifics, and direct claims survive extraction. Marketing language does not.
Get onto the aggregate pages. Directory and comparison pages get retrieved for exactly the queries where a buyer is choosing between options, and being listed on them puts you in answers you'd never reach with your own site alone.

That last one is the connection between this and the more conventional work of getting listed, which we go into in do directory backlinks help you get cited by AI.

A caution

Do not over-fit to any of this. These systems change, and a page engineered around a specific retrieval quirk will age badly. The durable version of this advice is much older than AI: write pages where the answer to a real question is stated clearly, in one place, in plain language. That worked before retrieval systems existed and it will work after these particular ones are replaced.

Frequently asked questions

How do AI answer engines decide what to cite?: Most run a search, fetch a handful of top results, split those pages into passages, pick the passages that best match the question, and attribute the sentences they used back to the page the passage came from.
Why does my page rank but never get cited?: Usually because the answer is spread across several paragraphs rather than stated in one. Retrieval works on passages, so an answer that only makes sense after reading the whole page rarely survives the chunking step.
Do I need to be on page one to be cited?: It helps a great deal, because most answer engines only fetch a small number of results. Being in the top handful for the exact phrasing of a question matters more than being first for a broad head term.