Structuring content so LLMs can quote it: a practical checklist

Why structure matters to an assistant

Search-grounded assistants like ChatGPT search, Perplexity, Gemini and Claude with web search share a broad pattern. They find candidate pages, fetch or look up their content, and generate an answer that cites some of those sources. None of the vendors publishes the full pipeline: how pages are chunked, how passages are scored, or why one source gets cited over another. We won’t pretend otherwise.

What we can reason about is the input. A page whose main content is readable without JavaScript, divided into clearly labeled sections, and written in self-contained passages is easier for any retrieval system to use. That holds whatever the ranking details are. The checklist below sticks to those controllable inputs.

Step 1: Put the content in the HTML response

Google renders JavaScript for Search. AI crawlers are a different story. Vendors generally don’t document JavaScript rendering for agents like GPTBot, OAI-SearchBot, ClaudeBot or PerplexityBot, so the safe assumption is that they read the initial HTML and nothing more. A client-only React or Vue app that ships an empty <div id="root"> may look blank to them.

Here’s what to check on your important templates:

1. Run curl -s https://example.com/page/ and search the output for a sentence from the main body copy. If it’s missing, crawlers that don’t render won’t see it either.
2. Load the page with JavaScript disabled in your browser. Headings, body text, prices and FAQs should all still be there.
3. If they aren’t, move those templates to server-side rendering, static generation or pre-rendering. Most modern frameworks support at least one of these.
4. Check that the crawlers you’ve allowed actually get a 200 response. Our AI crawlers and robots.txt playbook covers the access side.

Step 2: Write passages that stand on their own

An assistant usually quotes or paraphrases a passage, not a whole page. Write each section so it would still be accurate and understandable if it were the only thing the reader saw.

Answer first

Phrase each <h2> or <h3> as the question or task it covers, then answer it directly in the first one or two sentences. Put the nuance, exceptions and background after the answer. This is the same inverted-pyramid habit that works for featured snippets and for human skimmers.

Name things explicitly

• Repeat the subject instead of leaning on pronouns across section boundaries. Write “the Pro plan includes SSO”, not “it includes SSO”.
• Remove backward references like “as noted above” or “see the previous section”. They break when a passage is extracted.
• Define a term the first time it appears in each section where it matters, in one plain sentence.
• Use one name per concept. If the product is “Team plan” in one place and “Business tier” in another, you’ve made two things out of one.

Be specific and checkable

Concrete statements with units, dates, limits and conditions are easier to quote accurately than vague claims. “Supports files up to 2 GB on paid plans” says more than “handles large files”. If a fact changes over time, say when it was last verified.

Step 3: Use structure that survives extraction

Semantic HTML carries meaning that a parser can use. Visual styling alone does not.

1. One <h1>, then a logical heading hierarchy. Don’t skip levels for styling, and don’t use bold paragraphs as fake headings.
2. Ordered lists for sequences, unordered lists for sets. Steps belong in an <ol>.
3. Real tables for comparisons, with <th> header cells. A comparison drawn as an image or built from styled divs loses its row and column relationships.
4. Keep key facts in text. Pricing in an image, specs in a PDF, or FAQs that load only after a click are all at risk of being missed.
5. Add stable anchor IDs to headings, so a section can be linked and referenced directly.

Practical tip: paste a page’s server HTML into a plain-text converter and read the result top to bottom. If the plain text is confusing, out of order or missing facts, a retrieval system is working with the same mess.

Step 4: Make provenance obvious

Assistants and the people reading their answers both benefit from knowing who wrote something and when. Show it on the page and repeat it in structured data.

• A visible byline linked to an author page that describes the person’s relevant experience.
• Visible published and updated dates. Change the updated date only when the content actually changes.
• Citations to primary sources for factual claims, linked inline.
• Article markup that matches what’s on the page, connected to your Organization entity.

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Article",
  "headline": "How to size a heat pump for a small home",
  "datePublished": "2026-01-15",
  "dateModified": "2026-03-02",
  "author": {
    "@type": "Person",
    "name": "Jane Doe",
    "url": "https://example.com/team/jane-doe/"
  },
  "publisher": { "@id": "https://example.com/#organization" }
}
</script>

Schema is not a known ranking signal for LLM citations, and no assistant vendor has documented how it uses this markup. It’s still worth doing, because it makes your claims about authorship and dates explicit and consistent. The publisher reference ties into the entity work in our entity SEO for AI search playbook.

What doesn’t help

• Hidden text or content shown only to bots. It’s cloaking under search engine policies, and it’s fragile.
• Stuffing pages with prompt-style phrasing. Write for a reader with a question. Odd phrasing aimed at a model tends to read badly to humans.
• Relying on llms.txt. It’s a community proposal, and major assistants haven’t documented support for it. Treat it as optional, not as a substitute for crawlable HTML.

Run the checklist, then measure

Apply the steps to your ten highest-value pages first: core service pages, pricing, comparisons and your most-linked guides. Then track whether citations and referrals change over the following weeks, using the approach in tracking AI referrals and citations. Expect noise. A single prompt run tells you very little.

If you want the checklist applied across your templates by people who’ll also fix the rendering issues, that’s what our implementation sprints are for. Book an audit call to scope it.