How to write a page so ChatGPT can quote it – structure and examples
A page gets quoted by ChatGPT when it answers one specific question in the first two or three sentences, splits the rest into short self-contained blocks under question-style headings, and backs claims with concrete details a model can lift verbatim. AI assistants do not cite whole pages – they extract passages, so each paragraph has to make sense on its own, without the paragraph above it. Everything else (markup, crawler access, clean HTML) exists to make that extraction possible.
Why does ChatGPT quote some pages and ignore others?
When ChatGPT answers with links, it usually runs a search, fetches a handful of pages, and builds the answer from fragments of them. The model is not reading your site end to end; it is scanning for the chunk that matches the user's question most directly.
That means two filters decide whether you are quoted. First, retrieval: your page has to appear in the underlying search results for that query. Second, extraction: once fetched, your page must contain a passage that answers the question in a form the model can reuse without rewriting it from scratch.
The exact ranking logic differs between ChatGPT, Perplexity, Gemini, Copilot and Google AI Overviews, and none of them publish it. What is consistent across all of them is the passage-level behaviour: clear, self-contained, factual blocks are quoted far more often than long narrative prose.
What does a quotable page structure look like?
Think of the page as a stack of independent answer blocks rather than an essay. A reliable layout:
- H1 that states the question or task in the words people use.
- Answer block: 40–80 words immediately under the H1, before any heading, answering the title directly.
- Context paragraph: why the answer is what it is, or when it does not apply.
- H2 sections, each one a question, each containing 2–5 short paragraphs or a list.
- A table wherever you compare options, tools, formats or prices.
- A procedure as a numbered list, one action per step.
- FAQ section with 3–5 extra questions that did not fit the main flow.
- Author, publication date and last-updated date, visible on the page.
The pattern is deliberately repetitive. Repetition is what lets a model find the right block in under a second of processing.
How should the first paragraph be written?
Write it as a complete answer that could be pasted into a chat window with no edits. Include the subject noun (not "it"), the direct answer, and one qualifying condition.
Weak:
In this article we will look at several approaches to the problem and explain why there is no single right answer.
Strong:
A B2B landing page should load its main content server-side in under 2.5 seconds. Pages that render product details only in JavaScript are often fetched by AI crawlers before the content exists, so the assistant cites a competitor instead.
The second version contains a claim, a number, a mechanism and a consequence. That is quotable material. The first version contains nothing a model can carry into an answer.
How do you phrase headings so an AI matches them to a question?
Use the user's question as the heading, not a clever label. "How much does a technical SEO audit cost?" works; "Pricing considerations" does not.
Three rules that help:
- One question per heading, never two joined with "and".
- Put the key noun in the heading, not in the sentence before it.
- Keep headings under about 10 words so they survive as a snippet title.
Hierarchy matters too. Use H2 for each distinct question and H3 only for sub-parts inside a long section. Skipping levels or styling text to look like a heading without using a real tag hides the structure from parsers.
How long should a paragraph be?
Two to four sentences, roughly 40–80 words. A paragraph that runs past 100 words usually contains two ideas, and a model will either take half of it or skip it.
Make each paragraph self-contained. Replace "this", "that approach" and "as mentioned above" with the actual noun. If you quote the paragraph out of context and it still makes sense, it is ready.
Which page elements get quoted most often?
| Element | Why assistants use it | How to format it |
|---|---|---|
| Definition sentence | Answers "what is X" directly | "X is a … that …" in one sentence, early on the page |
| Numbered procedure | Maps to "how do I …" questions | One action per step, verb first, 5–9 steps |
| Comparison table | Compact, unambiguous, easy to summarise | Real HTML/Markdown table, short cells, no merged rows |
| Numbers with a source | Models prefer specifics over adjectives | Figure + unit + date + where it came from |
| FAQ block | Pre-formed question–answer pairs | Question as heading, answer in 2–3 sentences |
| Checklist | Easy to reproduce as a bullet list | Short imperative lines |
Avoid putting any of this inside images, PDFs embedded in iframes, or tabs and accordions that only build on click. If the text is not in the served HTML, assume it does not exist.
What does a rewrite look like in practice?
Before:
Pricing is something many clients ask about. There are a lot of factors involved and every project is different, so it is difficult to give a universal answer, but we always try to be transparent and find a solution that works for everyone.
After:
How much does a technical SEO audit cost? A technical SEO audit for a site under 500 pages typically costs [check: price range] and takes 5–10 working days. Price rises with the number of indexable URLs, the number of languages, and whether the site is server-rendered or a JavaScript SPA. Audits of ecommerce catalogues over 10,000 URLs are usually quoted individually.
The second version is longer, but every clause is extractable: a price, a timeline, three cost drivers and an exception.
What markup helps AI understand the page?
Structured data does not force a citation, but it removes ambiguity about what the page is and who wrote it. Three types cover most content pages.
FAQ markup, matching the visible FAQ section word for word:
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [{
"@type": "Question",
"name": "How long should a quotable paragraph be?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Two to four sentences, roughly 40 to 80 words, each understandable without the paragraph before it."
}
}]
}
Article markup with authorship and dates:
{
"@context": "https://schema.org",
"@type": "Article",
"headline": "How to write a page so ChatGPT can quote it",
"datePublished": "2025-01-15",
"dateModified": "2025-06-02",
"author": {
"@type": "Person",
"name": "Author Name",
"url": "https://example.com/team/author-name"
},
"publisher": {
"@type": "Organization",
"name": "Example",
"url": "https://example.com"
}
}
And a meta description that is itself a usable answer, not a teaser:
<title>How much does a technical SEO audit cost? | Example</title>
<meta name="description" content="A technical SEO audit for a site under 500 pages costs [check: range] and takes 5-10 working days. Price depends on URL count, languages and rendering.">
Never put a claim in schema that is not visible on the page. Mismatched markup is treated as spam by search engines and gives assistants a contradictory signal.
How do you make sure AI crawlers can actually read the page?
Quotability starts with access. Check three things.
- Server-rendered HTML. Fetch your URL with
curland search for your answer paragraph. If it is absent, the content depends on client-side JavaScript, and several AI crawlers will not see it. - Crawler permissions in robots.txt. Different bots serve different purposes, and their names change over time, so verify current names in each vendor's documentation.
User-agent: GPTBot
Allow: /
User-agent: OAI-SearchBot
Allow: /
User-agent: ChatGPT-User
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: Google-Extended
Allow: /
Sitemap: https://example.com/sitemap.xml
- No blocking at the edge. WAF rules, bot-protection services and country blocks often reject AI crawlers before robots.txt is even read. Check server logs for the user agents above.
Some publishers also add an llms.txt file – a plain-text map of the site's key pages for language models. It is a community proposal, not an official standard, and major AI vendors have not confirmed they consume it, so treat it as low-cost and optional:
# Example
> B2B analytics platform for logistics companies.
## Core pages
- [Pricing](https://example.com/pricing): plans, limits, billing terms
- [How it works](https://example.com/product): data sources and setup steps
- [Docs](https://example.com/docs): API reference and integration guides
What should you check before publishing?
- The first paragraph answers the title with no heading above it.
- Every H2 is a question someone would type.
- No paragraph exceeds four sentences or opens with an unexplained pronoun.
- At least one table, one numbered procedure and one FAQ block.
- Every number has a unit, a date and a source.
- Author and last-updated date visible in the HTML.
- Key content present in the raw HTML response, not just after JavaScript runs.
- Schema matches the visible text exactly.
- AI crawlers are not blocked by robots.txt or the firewall.
FAQ
Does structured data make ChatGPT cite my page?
No. Schema.org markup clarifies what the page contains and who published it, but citation depends on whether the page is retrieved for the query and whether it holds a passage that answers it. Treat markup as a supporting signal, not a lever.
Do I need separate pages for AI search and classic SEO?
No. The same qualities – clear structure, direct answers, factual specifics, fast server-rendered HTML – help in both. The main difference is emphasis: AI answers reward passage-level clarity, while classic search still rewards depth and internal linking across the whole page.
How many words should a quotable page have?
There is no fixed number. Cover the question completely and stop. A 700-word page with four well-formed answer blocks is quoted more often than a 3,000-word page where the answer is buried in paragraph nine.
Should I write an FAQ section on every page?
Only where real questions remain after the main content. A forced FAQ repeating what the body already said adds duplicate passages and dilutes the page. Three to five genuine follow-up questions is a good ceiling.
How do I know if an AI assistant is already citing me?
Ask the assistants your target questions directly and record which domains appear, then repeat monthly. Server logs showing AI crawler user agents, and referral traffic from chat domains, give a second, independent signal.
If you want to see how your own pages look to AI answer engines right now, run a free check in Zgadka: it scores GEO and SEO visibility, flags what blocks citation, and returns a report with fixes and ready-made files such as robots.txt, llms.txt and schema.org markup.