How to write pages AI search can quote

A retriever does not read your page. It cuts it into passages, scores those passages against a question, and puts the best few in front of a model. If the sentences that answer the question only make sense in the middle of your page, they are thrown away.

Updated 30 October 2026

What an answer engine actually does with your page

The steps are the same everywhere, even though the details differ per vendor: fetch the page, extract the text, split it into passages — usually at headings and paragraph boundaries — and rank each passage against the question. A handful of passages go into the prompt, and the answer is written from those.

Rank is computed per passage. That is the whole game. A page can be excellent, rank first in Google, and still contribute nothing to an AI answer because every paragraph that would have answered the question refers to something outside it.

1. Name the subject inside the passage

The passage is read on its own. “It”, “this plan” and “the product” are resolved against the page, which the model does not have in front of it.

The fix costs nothing and takes seconds: repeat the name. Repetition reads badly to a human editor and perfectly to a retriever.

2. Cut the dependencies

Anything that points outside the paragraph — “as mentioned above”, “the previous step”, “this feature”, “compared with option B” — is a reason to drop the passage. So is an answer that continues in the next paragraph.

The test: cover the rest of the page and read the passage aloud. If someone who walked in halfway through would not understand it, neither will the retriever.

3. Carry one thing worth quoting

Models quote specifics: a number, a date, a version, a price, a named comparison, an observation only you could have made. “Fast and reliable” is not quotable; “a median of 38 ms per request across 12 regions in March 2026” is.

Do not invent the figure to satisfy this rule. A page with no numbers can still be quotable through a specific, checkable claim about how something works — but if the answer to a question is a figure, the answer engine will quote a page that has one.

4. Let the structure do the splitting

• Use headings that name a question or a subject, not “Overview” and “More information”.
• Keep one idea per section; two ideas in one section means one of them is buried.
• Keep paragraphs short. One claim per paragraph is easier to retrieve than one claim in four.
• Use a table for a comparison. Tables survive extraction as rows, and a row is a complete passage.

5. Keep the text in the HTML

Most AI crawlers do not run JavaScript. If the text arrives after a client-side fetch, they see an empty shell — which is why the same page can rank in Google and be invisible to a retrieval bot that renders nothing.

Text inside images is not text either. Neither is a specification that lives in a PDF with no HTML summary. Server-render the content, or at least make sure the crawler gets the text in the first response.

What does not help

• Keyword repetition. Retrieval is semantic; a paragraph repeating a phrase twenty times ranks no higher.
• Structured data that claims things the page does not show. Mismatched markup is ignored, and repeated mismatches are a liability.
• “AI-optimised” sentence templates. The model is looking for an answer, not a format.
• Publishing more pages instead of making the ones you have quotable. Volume is not a passage.

A checklist for one page

• Every section names its subject without relying on a previous paragraph.
• No “as mentioned above”, no unresolved pronouns at the start of a passage.
• At least one specific, checkable fact per section.
• Headings are questions or subjects, not labels.
• Comparisons are tables, not prose.
• The main text is in the HTML, visible without JavaScript.
• Each passage still reads correctly when the rest of the page is hidden.