How do you write a page an AI engine can actually quote?

Summary

An answer engine retrieves a chunk of your page, not the page, and that chunk arrives with no heading above it and no paragraph before it. So every section has to answer its own heading in the first sentence, resolve its own pronouns, define its own acronyms, name its sources inline and run roughly 40 to 150 words. The 2024 KDD paper by Pranjal Aggarwal and colleagues at IIT Delhi and Princeton tested nine methods across a 10,000-query benchmark and found adding quotations, adding statistics and citing sources produced a 30 to 40% relative improvement on its Position-Adjusted Word Count metric — provided the quotations and figures are real.

Published

By making every section survive being read entirely on its own. An answer engine does not retrieve your page — it retrieves a chunk of it, drops that chunk into a context window alongside chunks from other websites, and writes an answer from the pile. Your section arrives with no title above it, no paragraph before it and no navigation around it. If it only makes sense in place, it makes no sense at all by the time it is being read.

That single constraint generates almost every rule below. This is the practical companion to how AI engines actually choose which sources to cite, which explains the retrieval mechanism this piece assumes.

What does the research say actually works?

The clearest published evidence is the paper that named the field. Pranjal Aggarwal and colleagues at IIT Delhi and Princeton University published GEO: Generative Engine Optimization at KDD 2024, testing nine optimisation methods across GEO-Bench, a benchmark of 10,000 queries. Its three best-performing methods were adding quotations, adding statistics and citing sources, reported as a relative improvement of 30–40% on its Position-Adjusted Word Count metric and 15–30% on its Subjective Impression metric. The authors also found effectiveness varied by domain, which argues against treating any of it as a universal recipe.

Two things that finding is not. It is not a revenue result, and it is not a promise for your category — it is a measurement of prominence inside a research benchmark. And it only holds when the quotations and statistics are real. Engines cross-reference claims, so a number that cannot be traced to a named source weakens the passage it was meant to strengthen.

What makes a passage retrievable?

Five properties. A section that has all five works as a standalone answer; a section missing any one of them tends to fail in a specific, predictable way.

Properties of a retrievable passage, and what breaks without each
Property What it means in practice Failure mode when missing
Answer-first The first sentence answers the heading directly, before any setup The retrieved chunk is preamble; the answer was in a paragraph that did not travel with it
Self-contained No pronoun refers to something outside the section "This approach also helps with…" — the engine has no idea what this is
Terms defined locally Acronyms expanded on first use in that section, not once at the top of the page The passage reads as jargon and matches fewer ways of asking the question
Sourced inline The source is named in the same visible sentence as the figure The number is unattributable, so a careful system declines to repeat it
Right length Roughly 40–150 words: long enough to be complete, short enough not to be split Too short and it says nothing; too long and it gets cut mid-argument

What does the difference look like on the page?

Take a section headed "Our approach to lead times". A version that fails retrieval:

As mentioned above, this is where our experience really shows. We have refined the process over many years and it delivers consistently strong results for our customers, which is why so many of them come back to us.

Every sentence depends on something outside itself. "This" has no referent. "Strong results" is unquantified. Retrieved alone, it answers nothing and could describe any company in any industry.

A version that survives retrieval:

Standard lead time for our servo drives is six weeks from order confirmation, or two weeks for the eleven part numbers held in stock in Rotterdam. Lead times are quoted at order rather than estimated, and a delay is notified within one working day of it becoming known. For custom-wound units the lead time is quoted individually, because it depends on core availability.

Same claim, but it is now specific, self-contained and attributable to a named company. It also happens to be the version a human buyer prefers, which is usually how you can tell the advice is not a trick.

How should headings be written?

As questions, phrased the way a buyer would type them. There are two reasons, and the second is the one people miss.

First, the heading is part of the chunk. A heading reading "Lead times" gives the retrieval step almost nothing; a heading reading "How long are your lead times?" matches the shape of the question directly.

Second, headings segment the page. Where you put a heading is where the page is likely to be divided, so headings are the mechanism by which you decide what a passage contains. A page with three headings and two thousand words has handed that decision to a chunking algorithm. A page with nine question-shaped headings has made it deliberately.

Does this mean writing for machines instead of people?

No, and it is worth being clear about why not, because "write for AI" is doing a lot of unearned work in marketing copy at the moment. Every rule above — answer the question first, do not make the reader hold context, define your terms, say where the number came from, keep sections short — is ordinary editorial discipline. It is what a technical buyer reading in a hurry wants anyway.

The genuine change is not the writing style. It is that the penalty for vagueness went up. Vague copy always failed with skim-readers; it now also fails a retrieval step that has no patience and no context at all.

Does any of this help if the page cannot be crawled?

None of it. Vercel and MERJ tracked over 500 million GPTBot requests in December 2024 and found no evidence of JavaScript ever being executed — GPTBot requested JavaScript files 11.5% of the time and ClaudeBot 23.8%, and neither ran them, with Gemini and Applebot the exceptions. If your pages assemble themselves in the browser, the best-written passage in your industry is invisible to most of these systems. Check that first; the writing is the second job, not the first.

What to change on your own pages this week

Pick your three most commercially important pages and do one pass on each: turn every heading into the question it answers, move the answer into the first sentence beneath it, resolve every pronoun that points outside its own section, and put a named source in the same sentence as every figure. That is a couple of hours of work and it is measurable — re-run the same buyer questions afterwards and see whether the passages that get cited change.

If you would rather see where you stand before rewriting anything, the diagnostic comes first. The service it belongs to is getting found when buyers ask AI.

Run a free AI visibility check

Frequently asked questions

What makes a passage retrievable by an AI engine?

Five properties: it answers its heading in the first sentence, it contains no pronoun referring to something outside the section, it expands its own acronyms rather than relying on a definition at the top of the page, it names the source of any figure in the same visible sentence, and it runs roughly 40 to 150 words so it is neither empty nor split mid-argument.

Does adding statistics and quotations to a page really improve AI visibility?

The 2024 KDD paper on generative engine optimisation, by Pranjal Aggarwal and colleagues at IIT Delhi and Princeton, tested nine methods across GEO-Bench, a benchmark of 10,000 queries, and found its three best performers were adding quotations, adding statistics and citing sources — a 30 to 40% relative improvement on its Position-Adjusted Word Count metric. The condition is that they are real and traceable, because engines cross-reference and an unattributable figure weakens the passage it was meant to strengthen.

Should headings be written as questions?

Yes, for two reasons. The heading travels with the chunk, so a heading phrased the way a buyer would type the question matches far better than a one-word label. And headings decide where the page is likely to be divided, so writing more of them is how you choose what each passage contains instead of leaving it to a chunking algorithm.

Is writing for AI different from writing for people?

Not in what it asks for. Answering first, resolving pronouns, defining terms and citing sources are ordinary editorial discipline and are what a technical buyer reading in a hurry wants anyway. What changed is the penalty for vagueness: vague copy always failed skim-readers, and it now also fails a retrieval step that has no context at all.

Find out what the engines currently say about you

Send me your company name and website. I run a set of buyer questions across ChatGPT, Perplexity, Gemini, Claude and Google's AI Overviews, and send back a recorded walkthrough of what came out: where you appeared, where you did not, who was named instead, and the two or three structural reasons why. No charge, no obligation, and you keep the findings whether or not you decide to hire me.

I run these myself, so there is a queue. Expect a few working days rather than an instant report.