How do I make a page quotable by AI?
The structural changes that make a passage extractable — most of them a rearrangement of what you have already written.
Put the answer first. A heading that asks the question, one or two sentences answering it directly underneath, then the explanation. Everything else — attribution, dates, structured data — is secondary to that one change.
The one structural change
Most writing builds towards its conclusion: context, then reasoning, then the answer. That is good prose and poor extraction, because the quotable sentence is at the bottom and everything above it is required to make sense of it.
Inverting it costs nothing in quality if the elaboration still follows:
- Heading: the question, phrased as someone would ask it.
- First one or two sentences: the answer, complete and self-contained.
- Then: the reasoning, the caveats, the examples, the depth.
A reader gets the answer immediately and can keep reading. An extraction system gets a passage that works alone. Nobody is worse off.
What makes a passage self-contained
| Not extractable | Extractable |
|---|---|
| "As mentioned above, this is why it fails." | "CSP fails when unsafe-inline is present, because it permits exactly what CSP exists to prevent." |
| "It depends on several factors." | "For sites under 500 pages, a sitemap adds little — crawlers find everything by following links." |
| "See the table below." | "HSTS should use max-age=31536000, which is one year." |
| "This one is my favourite." | "Referrer-Policy: strict-origin-when-cross-origin prevents full URLs leaking to third parties." |
The right column shares one property: each sentence names its own subject and would make sense pasted into a conversation with no surrounding text.
Structures that extract well
- Question headings with direct answers. The strongest single pattern.
- Definitions stated plainly. "X is Y" — a complete claim in one sentence.
- Numbered procedures. Self-contained by construction, and each step names its own action.
- Comparison tables. Extractable when each row makes a complete statement, less so when cells are fragments requiring the header to parse.
- Specific numbers with their units and context. "One year" beats "a long time"; "31536000 seconds, which is one year" beats both.
Structures that extract badly
- Anything behind an interaction. Accordions, tabs and modals that build content in JavaScript are invisible to non-rendering crawlers. If the content is in the HTML and merely hidden by CSS, it is generally fine.
- Answers in images. A diagram containing the answer with no text equivalent is unreadable to everything.
- Pronoun chains. "It does this because it depends on that" — nothing in the sentence identifies its own subject.
- Answers deferred. "We will come back to this later" removes the sentence anyone would quote.
The supporting details
Once the structure is right, these help a system cite you responsibly rather than merely find you:
- A named author with a real identity, not "admin" or "the team".
- An organisation, ideally marked up so it is unambiguous.
- A date, and an honest one — a modified date that updates without the content changing is worse than none.
- Structured data describing what the page is, which removes ambiguity rather than adding ranking.
- Original detail somewhere: a number, a measurement, an observation nobody else has published.
What not to do
Do not strip depth to make a page more extractable. A page reduced to disconnected one-line answers is worse for readers and gains nothing — extraction rewards a good answer near the top, not the absence of everything after it. Lead with the answer, then write the rest properly.
What our audit reports about this
Every item below is measured directly, not inferred. Run it against your own site and the result names the exact rule or header responsible.
- Whether headings are question-shaped and whether a direct answer follows each one.
- Whether content sits inside interactive components that non-rendering crawlers cannot reach.
- Whether author, organisation and dates are present and machine-readable.
- Whether structured data is present and valid, giving an unambiguous account of what the page is.
For agents and scripts, the same measurement is at
/api/v1/content?url=yoursite.com —
see the API documentation.
Related questions
Does this hurt readability?
It generally improves it. Leading with the answer is standard practice in technical writing and journalism for good reason — readers who want the answer get it, and readers who want the reasoning read on.
Should every page have an FAQ section?
Only where there are genuine follow-up questions. A real FAQ is close to ideal extraction structure. An invented one is padding, and padding is detectable.
Do accordions hurt?
It depends on the implementation. Content present in the HTML and hidden with CSS is generally readable. Content that a script fetches or builds when the section is opened does not exist for a non-rendering crawler at all.
How long should the direct answer be?
One or two sentences. Long enough to be complete and self-contained, short enough to be quoted whole. If it runs to a paragraph, the question is probably several questions.
Read next
How do answer engines choose between two similar pages?
When two pages cover the same ground, the tiebreakers are extractability, specificity, corroboration and attribution — in roughly that order.
ReadHow do I write content that AI assistants will quote?
Assistants extract passages, not pages. Everything follows from that one fact.
ReadHow do I add an FAQ section that search engines actually use?
Real questions, visible answers, correct markup — in that order. Schema on questions nobody asked achieves nothing.
ReadDoes schema markup help with AI search?
What structured data actually does for answer engines — and why it is worth adding for a reason other than the one usually given.
Read