Writing Pages That AI Answers Actually Cite

An assistant answering a question does not read your page. It reads a fragment of it (a few hundred words, retrieved out of context, sitting in a list alongside fragments from other sites) and decides whether that fragment is worth quoting.

Everything about writing for AI answers follows from that one mechanic. Your page is not the unit of competition. The passage is. A brilliant article whose key claim only makes sense after three paragraphs of setup will lose to a mediocre one that answers the question in its first sentence, because the retrieval system will never show the model those three paragraphs.

How retrieval actually decides

Simplified, but accurate enough to write against:

The step that matters is the split. Wherever your page gets cut, the fragment on either side of the cut has to stand on its own. "As we saw above, this means the second approach is usually better" is a perfectly good English sentence and a worthless retrieval passage: it names neither this, nor the second approach, nor what it is better at.

The five properties of a citable passage

1. It answers before it explains. Traditional writing builds to a conclusion. Retrieval truncates, so the conclusion goes first and the reasoning follows. This is closer to journalism's inverted pyramid than to essay structure, and it costs you nothing with human readers, who also prefer the answer up front.

2. It is self-contained. Every passage restates its subject. No naked pronouns pointing at earlier sections, no "as mentioned above", no "the tool" where you could write the product's name.

3. It is specific and attributable. "Significantly faster" is unquotable. A sentence of the form "reduced median response time from 840 ms to 210 ms, measured across 10,000 requests in March 2026" is quotable, because a model can carry the claim into an answer without inventing anything, and the number is checkable. Vague claims get paraphrased into generic filler and lose the citation to whoever was specific.

4. It is structured where structure is honest. Tables, ordered lists and short definitions survive extraction cleanly. A comparison written as three long paragraphs has to be re-derived by the model; the same comparison as a table can be lifted intact.

5. It is dated. Facts about a fast-moving field are worth more when the reader can see when they were true. Publication and revision dates, and dates inside the claims themselves, all help a model prefer your version over an undated competitor's.

Rewriting a passage

The difference is small on the page and large in retrieval.

Before

As we discussed in the previous section, this can be a significant problem for many organisations. Our research shows that the impact is considerable, and most teams are surprised by the results when they finally look into it.

Nothing here survives extraction. No subject, no number, no date, no source. Retrieved on its own it answers no question at all.

After

Most AI-referred visits arrive with no referrer header (a consequence of referrer policies and of assistants that are read rather than clicked), so standard analytics files them under Direct. Cloudflare's 2025 reporting put some operators' crawl-to-refer ratio in the tens of thousands of pages per visitor returned, which is why a site's reported "AI traffic" is usually a fraction of its real AI exposure.

Same length. Names its subject, carries two claims that are each attributed (one to a named source, one to its mechanism) and answers a question, why does AI traffic not show up in analytics, without any surrounding context.

Headings are queries

Retrieval is driven by what someone typed. Headings that mirror those phrasings both improve matching and force the section below to be a direct answer.

Two mechanical rules: one H1 per page, and a heading hierarchy that does not skip levels. Heading structure is frequently the boundary along which pages get chunked, so a broken hierarchy produces badly-cut passages before anything semantic is even considered.

The things that disqualify you before the writing matters

Content that requires JavaScript. Most AI crawlers fetch HTML and never execute scripts. If your content is rendered client-side, there is nothing in the response for them to read, and no amount of passage craft fixes an empty document. Server-render anything you want cited.

Being blocked. Check robots.txt for the search-index crawlers specifically. Blocking OAI-SearchBot or PerplexityBot (usually by inheriting somebody's blocklist) removes you from consideration entirely, which is a decision worth making deliberately rather than by copy-paste.

Being slow. Crawlers have timeouts. Consistently slow pages get fetched less often and sometimes not at all.

Never being crawled. A page that no search-index crawler has fetched cannot be cited regardless of quality. Crawl coverage is the precondition for everything in this post, and it is measurable server-side today.

Structured data, honestly

Schema.org markup is worth adding, with realistic expectations. It does not force a citation. What it does is remove ambiguity (who published this, when, what it is about, what the questions and answers on the page are) for systems that parse it.

Three types earn their keep:

• Article with datePublished, dateModified and a real author. Cheap, unambiguous, and the dates are the part that matters.
• FAQPage on pages with genuine question-and-answer sections. Each Q&A becomes an explicitly labelled, self-contained unit, which is the same property retrieval rewards, expressed in markup.
• Organization with sameAs links, so the entity behind the content is identifiable rather than inferred.

Generate it from the same source as the visible content. Two hand-maintained copies of the same FAQ will drift within a quarter, and a markup block that contradicts the page is worse than none.

Measuring whether any of it worked

The failure mode of every AEO checklist is that it produces activity with no feedback loop. Three measurements make it a loop, and all three are computable from your own systems:

The middle one requires building a prompt set (twenty to fifty questions a real buyer would ask) and checking them on a schedule. It is manual, it is genuinely tedious, and it is the only direct evidence of citation that exists, since no assistant reports impressions to publishers. Keep the prompt list fixed so the series stays comparable, and record the date each time.

Then remember the lag. Content changes take weeks to propagate through crawling and indexing, so a two-week readout is noise. Measure your own crawl-to-citation interval and set the review window from it.