What 'prompt extraction' means
Prompt extraction is the process by which an LLM selects which spans of text from its retrieved candidate sources will be used to construct the final answer it returns to the user. Sometimes those spans are quoted verbatim with attribution. Sometimes they're paraphrased into the model's voice. Sometimes they're combined with spans from other sources into a synthesis. The pattern depends on the model, the system prompt that frames the retrieval, and — crucially — the structure of the source page itself.
For the purposes of AEO, what matters is that you can substantially influence which of your sentences get extracted by writing them in particular ways. The patterns below come from sampling tens of thousands of cited answers across the major engines through Q1 2026.
The 6 extraction patterns we observe most often
Model finds a self-contained sentence that directly answers the user's question and quotes it. Most common for definitional and procedural queries.
Numbered or bulleted lists get extracted as a unit, frequently with their original ordering preserved. Procedural content benefits disproportionately.
Comparison tables get parsed and individual rows surfaced as bullet points or as inline phrases. <table> markup matters here; styled divs don't extract reliably.
Model takes one sentence each from 3–5 sources and stitches them. Each source gets a brief citation but no single one is quoted at length.
Model extracts a definition from one source and an example from another. Pages that provide both adjacent get cited disproportionately.
When the model wants to nuance a claim, it pulls explicitly contrarian sentences. Pages that surface counter-examples cleanly get cited as the nuancing source.
How extraction differs by model
| Engine | Dominant pattern | Typical span length | Citation style |
|---|---|---|---|
| Perplexity | Verbatim short quote | 20–60 words | Inline numbered footnote |
| ChatGPT browsing | Synthesized paraphrase | Variable | End-of-paragraph footnotes |
| Claude web search | Long extracted blocks | 60–200 words | Full inline attribution |
| Gemini | Heavily synthesized | Short or none verbatim | Inline 'According to' phrasing |
| Google AI Overview | Mixed extract + paraphrase | Variable | Sidebar source cards |
| Copilot | Numbered footnotes | 30–80 words | Inline numbered footnotes |
The practical implication: optimizing for verbatim extraction (Perplexity / Claude) and optimizing for paraphrase-with-citation (Gemini / ChatGPT) require slightly different page craft. The intersection — which is what we recommend most clients aim for — is short, declarative, self-contained sentences in answer blocks, paired with denser explanatory paragraphs below.
How to write for extraction (without writing badly)
The mistake most people make when 'writing for AEO' is mechanically front-loading every page with bullet-list answer blocks at the expense of readability. The pages that win extract well and read well. Six rules:
- Lead with one declarative sentence under the H1< 250 chars, no hedging, no marketing copy. This is the highest-leverage extraction surface on the page.
- Make every H2 a question or a clear noun phraseHeadings are the model's primary signal for what the section is about. Vague 'Overview' or 'Strategy' headings extract poorly.
- Use one claim per sentence in the first paragraph after each H2Compound sentences fragment poorly during extraction. The first paragraph carries most of the extraction weight.
- Use real semantic HTML for lists and tables<ul>, <ol>, <table>. Styled divs render visually but extract unreliably.
- Pair definitions with concrete examples in the same sectionThe 'definition + example' extraction pattern is one of the strongest. Make it easy by colocating them.
- Surface the contrarian view explicitlyIf your topic has counter-arguments, address them in their own section with a clear 'However...' framing. Models pull these aggressively.
Why some pages get quoted verbatim and most don't
Three properties predict verbatim extraction more than anything else we've measured:
If a single sentence on your page would make a coherent, accurate answer to the user's question, the model has every incentive to quote that sentence and cite you. If the answer requires the model to assemble three half-sentences from across your page, you'll get paraphrased — and cited less.
- Self-containment. The candidate sentence makes sense on its own without the rest of the page.
- Declarative form. No questions, no hedging, no 'we believe.' Direct claims.
- Surrounding delimitation. The sentence is easy to delimit — at the start of a paragraph, after an H3, in a list item, or in a callout.
When you put a sentence at the start of a paragraph immediately under an H2 that asks the question that sentence answers, you're doing roughly half the model's work for it. That's why direct-answer formatting earns the citation lift it does.
Where extraction is heading next
Two near-term shifts to design for:
- Tighter source attribution standardsMajor engines are converging on more precise citation — sentence-level rather than page-level. Pages with anchor-able sentences (#section IDs, schema.org/cite annotations) will benefit.
- Increased use of structured data for extractionFAQPage and HowTo schema are already heavily used. We expect Article + Speakable + Claim schema to see increased weight, particularly for citation-quality assessment.
- Engine-specific source preferencesEach engine is developing distinct trust profiles — Perplexity weights primary research heavily, Gemini weights long-tenure domains. Expect more divergence.
- Real-time citation rankingLive-mode extraction is becoming nearly real-time; pages updated today can earn citations within hours. Stale answer blocks lose to fresher rewrites quickly.
The honest summary
Prompt extraction is not magic. It's a predictable process of selecting self-contained, declarative, well-delimited spans from candidate sources. Pages that make extraction easy get quoted; pages that don't get paraphrased or ignored. The on-page changes that earn extraction are also the on-page changes that earn human readers — which is why this discipline is converging with classic content design rather than diverging from it.
Frequently Asked Questions
The questions we get when explaining how LLMs select what to quote.