A guide to choosing a heat pump sits on the first page of Google for its keyword, and the AI Overview above it cites two other pages instead. Open those two and the pattern is hard to miss: each one states the answer in a sentence you could lift out and paste into a chat window. Assistants quote passages, not pages, so the job is to write passages that survive being quoted on their own.
Put the answer first, then the reasoning
When an assistant retrieves your page it's looking at chunks, roughly a heading plus the paragraphs under it, and it scores each chunk by how well it matches the sub-question being answered. If the heading asks the question and the first sentence answers it, the chunk scores well and the model has a ready-made sentence to cite. If the answer arrives after three paragraphs of background, the chunk looks like background.
The pattern is simple to apply. Phrase the heading the way a person asks. Answer in the first sentence, plainly. Then give the reason or the method. Then the exceptions. A store selling trail shoes writing a section called "Do trail shoes work on pavement" would open with "Trail shoes work on pavement for short stretches, but soft lugs wear down quickly on tarmac and the ride feels harsher than a road shoe", and only then explain why the lugs matter.
Journalists call this the inverted pyramid. It was built for readers who stop early, which is exactly what a retrieval system does.
Make every section stand on its own
A chunk gets separated from its neighbours, so anything that depends on the paragraph before it breaks. "This makes it faster" loses its meaning when "this" lived in the previous section. "As mentioned above" points at nothing. A section that opens with a pronoun is a passage the model can't safely quote.
A few habits fix this:
- Name the subject in the first sentence of every section, even when it feels repetitive.
- Keep each section to one idea, so the chunk is coherent from its first line to its last.
- Use a table when items are parallel (a comparison of three plans) and a numbered list when they are sequential (the steps to claim a warranty), because lists and tables survive extraction intact.
- Put the key fact in text, never only in an image or a chart, since most AI crawlers read the HTML and nothing else.
The free Answer engine optimization tool counts question-style headings, lists and tables for this reason. They're the structural signs that a page is made of quotable units rather than one long run of prose.
Make claims specific, and say where they come from
Assistants are built to attach a source to a statement, so a specific, attributable sentence is a stronger candidate than a vague one. "Most customers see results quickly" gives the model nothing to hold. "A standard boiler service covers the burner, heat exchanger, flue and seals, and includes a gas pressure check" gives it a fact with edges.
That means naming the source of any figure in the same sentence, dating anything that changes, and showing who wrote the page and why they'd know. The E-E-A-T lesson covers what those trust signals look like on a page. It also means leading with the facts only you have: your pricing, your process, what you've seen across your own customers. Every competitor can paraphrase a general explainer. Only you can state what your own business does, and that is the kind of claim an assistant can't get anywhere else.
Cut the fog while you're there. Hedged openings, keyword-stuffed introductions and paragraphs that promise an answer without giving one are skipped by the model for the same reason a reader skips them.
Structured data helps at the margins. Article, Organization and Product markup remove ambiguity about what the page is and who published it. FAQPage markup is still valid Schema.org, but Google's Search Central updates record that FAQ rich results have been retired, so add it only where the questions are genuinely on the page and don't expect a search feature from it.
Cover the sub-questions as well as the headline
The previous lesson showed that one question fans out into several searches. The page that gets cited most often is the one that answers several of those sub-questions, each in its own clean section, rather than the one that answers the headline and stops.
Run the query through the free AI query fan-out simulator and hold your outline against the list. Each sub-question that belongs on this page gets a section with the answer-first structure above. Each one that doesn't belong becomes its own page, linked from this one, which is how a topic cluster forms. Don't force everything onto one page: a page with four tight answers beats a page with twelve vague ones, and it reads better to the human who arrives from the citation.
What to take away
- Assistants quote chunks of a page, so each section needs a question-shaped heading and a first sentence that answers it outright.
- A section must make sense when lifted out alone: name the subject, keep to one idea, and use lists and tables for parallel facts.
- Specific, sourced, first-hand claims get cited; hedged marketing prose gets skipped.
- Cover the sub-questions an assistant researches, each in its own section or its own page, instead of answering only the headline query.
Next
Your passages are ready to be quoted, so make sure the crawlers that feed the assistants can reach them in llms.txt and AI crawlers.