Passage-Level Optimisation: How AI Models Chunk Your Content Before Citing It

Passage-level retrieval is the mechanism by which AI engines decide what to cite from your website. When ChatGPT, Perplexity, or Claude retrieves content from the web, it does not read your full page and decide whether to recommend it. It splits your page into segments, evaluates each segment independently, and cites the segments that best match the query. If your page contains one excellent answer buried in three paragraphs of preamble, the excellent answer may never be extracted. Understanding this mechanism changes the entire approach to AEO.

Research into AI citation patterns shows that 44% of all citations come from the first 30% of a page's text. The implication is not just that you should front-load your best content - it is that the structural decisions you make at the paragraph and heading level directly determine which parts of your page AI models can find, extract, and quote independently of the rest.

What passage-level retrieval actually means

When an AI engine using retrieval-augmented generation (RAG) responds to a query, it first runs a search of the web or its index, retrieves candidate pages, and then processes those pages through a chunking step. Chunking splits each page into segments of roughly 100 to 300 words at structural boundaries. Each resulting chunk is converted into a semantic vector and scored for relevance to the query. The highest-scoring chunks are pulled into the context window that the model uses to compose its answer.

This is the core difference between SEO and AEO: in SEO, you optimise a page to win a ranking position. In AEO, you optimise individual passages to win a relevance score in a local competition against every other chunk in the retrieval pool. A competitor's 200-word, self-contained answer on a lower-authority domain may beat your 3,000-word article if your article never produces a clean, directly relevant chunk.

How AI models chunk your content

Chunking algorithms vary between AI systems, but the structural signals they use are consistent. Content is split at heading boundaries, paragraph breaks, and list delimiters. This means your heading hierarchy and paragraph structure directly control which ideas end up in the same chunk versus in separate, independently scored chunks.

• Each H2 and H3 heading typically starts a new chunk - the heading text is usually included at the top of the following chunk as context
• Each paragraph is a potential chunk unit if it contains a coherent, self-standing idea
• List items within a ul or ol are usually grouped into a single chunk, which makes lists highly extractable as a unit
• A paragraph over 200 words may be split mid-sentence by the chunker, producing chunks with no clear topic boundaries
• Code blocks and structured markup are typically extracted as standalone, typed chunks

The practical implication is significant: every H2 section of your page is a discrete citation opportunity. If the opening sentence of that section does not state the answer clearly, the chunk starting there will score poorly for any query about that topic, even if a strong answer appears two paragraphs later in the same section.

The position bias: why your first 30% dominates citation

A consistent finding across analysis of ChatGPT, Perplexity, and Gemini citation behaviour is that nearly half of all AI citations come from content located in the first 30% of a page. Two factors drive this. First, RAG architectures often apply a mild positional weighting that slightly favours earlier chunks when relevance scores are similar. Second, and more importantly, pages that bury their key points produce a concentration of low-relevance chunks near the top and high-relevance chunks near the bottom, inverting the distribution that retrieval systems expect.

The structural fix is to write using an inverted pyramid for every section, not just the page overall. State the most important answer in the first sentence under each H2 heading. Use subsequent paragraphs to expand, qualify, and provide evidence. This approach produces high-relevance chunks at the top of every section, not just at the top of the page.

Writing a self-contained passage

A self-contained passage is one that delivers a complete, useful answer without requiring the reader to have seen the surrounding context. AI retrieval systems favour self-contained passages because they can be inserted into a generated answer without the model needing to add a preamble or caveat explaining what the passage is referring to.

1. State the answer or key claim in the first sentence, not the second or third
2. Define any technical term on its first appearance within the same sentence where it appears
3. Keep the passage under 150 words so it fits cleanly as a citation without editing
4. Write so that a reader who arrives at this passage with no prior context would understand its full point
5. End on a complete thought, not a continuation that requires the next paragraph to make sense

The self-containment test is simple: copy a single paragraph and paste it into a blank document. If a reader who had not seen your heading or preceding paragraphs would fully understand the point, the passage is extraction-ready. If they would need to ask "what is this referring to?", the passage starts too late in the thought and should be rewritten so the referent appears in the first sentence.

Structural changes that improve passage extraction

Use headings as chunk anchors

Each H2 and H3 heading signals the start of a new topic to the chunking algorithm. The heading text is typically included at the start of the chunk that follows it, giving the chunk its primary topic signal. For this reason, descriptive, direct headings consistently outperform clever or abstract ones for AEO. Headings phrased as questions perform especially well because they directly match the conversational query format that AI retrieval systems score against.

Keep paragraphs under 120 words

Long paragraphs are the most common structural cause of poor passage extraction. A paragraph of 300 words typically contains two or three distinct ideas. The chunking algorithm will either split it mid-idea, producing ambiguous chunks, or treat it as a single unfocused unit that scores weakly for any specific query. The 120-word target aligns with the chunk size used by many RAG implementations for expository content. Paragraphs that stay consistently under this limit produce cleaner extraction and higher relevance scores.

Lists are highly extractable units

Unordered and ordered lists are processed as coherent units by most chunking algorithms, and structured data formats are cited roughly three times more often than equivalent paragraph-only content. A list of 5 to 7 specific, concrete items on a well-defined topic produces a single chunk that the retrieval system can easily match to queries like "what are the factors that affect X?" or "what are the steps to Y?". Each list item should be specific enough to stand alone as a fragment of a useful answer.

This is not a minor edge case: a 2026 meta-analysis of roughly 400 million AI citations found that list-format pages accounted for 63% of them, with the numbered listicle alone making up close to 22% of all website citations. See our guide to why listicles get cited more than prose for how to structure a ranked list a model will actually extract, and when a definitional paragraph still beats it.

What passage-level optimisation is not

Passage-level optimisation is not about placing target keywords at the start of every paragraph or writing every section as an isolated definition. Chunking algorithms operate semantically rather than lexically - they embed each chunk as a vector and retrieve based on vector similarity to the query, not keyword overlap. Writing that mechanically places query phrases at the top of every paragraph is detectable as low-value content and scores below naturally written, coherent passages on the same topic.

• Do not sacrifice natural flow between paragraphs - passages that read as part of a coherent article are more trusted than isolated atomic fragments
• Do not repeat the same phrasing at the top of every paragraph; variety signals genuine content depth
• Do not strip all context from your opening paragraph in pursuit of a clean snippet - a genuinely informative introduction that leads with the core answer serves both human readers and passage extraction
• Do not optimise only the first few sections - every section is a citation opportunity, and a strong second half of the page compounds the overall citation surface

Testing passage quality before publishing

Before publishing or updating a key page, apply the passage test to every section. Take each H2 block, including its heading and the paragraphs beneath it, and evaluate it in isolation. Does it answer the question implied by its heading? Does every paragraph open with a clear claim? Would a reader who arrived at this section via a direct anchor link understand the full point without scrolling back? Pages that pass this test for every section are well-positioned in the passage extraction competition.