How to Structure Content So AI Will Quote It

What structure actually means here

When people ask how to structure content for ai, they usually mean formatting. More bullets. More bold. Break up the walls of text. Most advice on how to optimize for llms stops right there.

That is not it.

Structure means your page is built out of pieces that survive on their own. A model almost never reads your page the way you do. It pulls a slice, a few hundred words at most, and uses that slice to write one sentence in an answer. If your slice only makes sense because of the three paragraphs above it, it is a bad slice, and you lose.

So the question is not whether the page looks tidy. It is whether any single section could be carried off on its own and still hold up.

Headings get you found. The body gets you quoted.

This is the most useful finding of the year and hardly anyone is talking about it.

In February 2026 a team led by Sunghwan Kim built SAGEO Arena, a test setup that runs a whole search-and-answer pipeline over a big collection of real web pages. Instead of testing one step, it tests all three: finding the page, ranking it, and writing the answer. The paper was updated in August 2026.

They tried rewriting just the body text of pages, which is what most ai content optimization advice tells you to do. Presence in the top 20 results fell from 0.58 to 0.53, about 9 percent worse. After reranking, presence in the top 10 fell from 1.00 to 0.84, about 16 percent worse. Their words: optimizing body text alone consistently degrades visibility across all stages.

Then they tried the opposite and rewrote only the structural parts, things like titles and headings. Retrieval went the other way, from 0.58 to 0.71, about 22 percent better. But the reranking number still dropped.

The reason the paper gives is simple. Structural text is packed with the words people actually type, so retrievers grab onto it. When it comes time to write the answer, though, the generator mostly reads the body.

You need both, and you need them doing different jobs.

What a heavily quoted page looks like

A second paper, from April 2026, measured the pages themselves. Zhang Kai, He Xinyue and Yao Jingang, who describe themselves as independent researchers in China, ran 602 controlled prompts across ChatGPT, Google and Perplexity and logged 21,143 citations. They pulled apart two things most people mash together: getting cited at all, and actually shaping what the answer says. They call the second one absorption.

Pages in the top quarter for influence had 11.44 times more words than pages in the bottom quarter, 12.50 times more headings, and 8.94 times the list density. Pages carrying definitions, statistics, comparisons or step-by-step instructions scored 41 to 77 percent higher.

Be careful with that. It describes what winning pages look like. It is not proof that adding headings to your page will make it win. Long pages naturally carry more headings, so the heading count may just be a shadow of the word count.

The FAQ fight nobody has settled

Here are two 2026 datasets pointing in opposite directions.

Discovered Labs, an agency that sells answer engine optimization and runs its own visibility tracker, so read this as vendor research, analysed 2 million citations across 10,000 cited pages over six months on ChatGPT, Claude, Google AI and Gemini. They controlled for domain and for content depth and ran nine robustness checks. FAQ sections came out as the strongest on-page signal they tested, at plus 0.07. A short summary block at the top was next, at plus 0.05. Schema markup, Lighthouse scores and Core Web Vitals showed no independent effect at all once depth and domain were accounted for.

Zhang's team found the reverse on FAQs. Question-and-answer formatting on its own scored 0.0947 mean influence against 0.1005 for everything else, which is 5.74 percent worse.

Nobody has reconciled the two. One possible reason is that they measure different things. Discovered Labs counted citations. Zhang's team measured how much a page shaped the answer text. A page could get picked up more often and still say less. That is a guess, not a finding.

The practical read: an FAQ block is cheap and it will not hurt your reader, so add one. Just do not build your plan on it.

Write each section so it can be carried away

Researchers at Wroclaw University of Science and Technology compared nine ways of cutting documents into pieces for retrieval, published on 30 May 2026. The methods that produced fewer, more coherent pieces beat the ones that chopped at fixed lengths, and not by a little. On one measure, recall at 10, the best coherent method hit 78.16 percent against 44.75 percent for fixed-size cutting.

Their tests ran on research benchmark collections rather than on ChatGPT or Google, so treat this as a hint about the machinery, not a measurement of your site. The hint is consistent though: where a piece starts and ends changes whether it gets found at all.

You do not control how an engine cuts your page. You do control whether the cut lands somewhere survivable. Open each section by naming its subject instead of saying this, it, or as we saw above. Put the short direct answer right under the heading. That little unit has a name and its own post, the answer capsule.

Match the page to the question being asked

If you only remember one thing about how to write content for ai search, make it this one. Discovered Labs' biggest number was not a formatting number at all. It was prompt-content alignment, at plus 0.37, roughly three times the next page-level signal. One standard deviation more alignment lined up with about 30 percent more citations.

In plain words: the page that wins is the page that is obviously about the exact thing being asked.

Our own study, 40 prompts run on ChatGPT and Google on 28 August 2026, points the same way. It is our own research, one run per prompt, so hold it loosely. Across eight B2B software categories we logged 490 cited URLs. On head-to-head questions, comparison pages took 56.7 percent of citations. On pricing questions, pricing pages took 53.4 percent. On broad category questions, listicles took 30.0 percent.

The engines also disagreed with each other. Product pages were 37.7 percent of ChatGPT's citations and 6.5 percent of Google AI Overview's. Listicles ran the other way, 5.3 percent against 16.2 percent.

No single page shape wins. The shape has to match the question.

What to do this week

Pick one page. Just one.

1. Write down the exact question that page should win.
2. Make the title and the H2s say that question in the words a person would type.
3. Under each H2, answer in the first two sentences.
4. Read each section on its own, with everything else covered up. Fix anything that leans on what came before.
5. Put one hard fact in each section: a number, a date, a price, a definition.

Do not rewrite the body and leave the headings alone. That is the exact move the SAGEO Arena test says makes things worse. If you want a second pair of eyes on which pages to start with, that is what our audit is for.

Sources

• Sunghwan Kim, Wooseok Jeong, Serin Kim, Sangam Lee, Dongha Lee. SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization. arXiv 2602.12187, v1 12 February 2026, v2 7 August 2026. Table 2.
• Zhang Kai, He Xinyue, Yao Jingang. From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms. arXiv 2604.25707, 29 April 2026. 602 prompts, 21,143 search-layer citations.
• Discovered Labs. What actually drives AI citations: a statistical analysis of 2M AI citations across 10K pages. Vendor research by an agency that sells answer engine optimization.
• Mateusz Smigielski, Michal Rajkowski, Mateusz Zbrocki, Michal Bernacki-Janson, Karol Kunicki, Julianna Godziszewska, Maciej Piasecki, Konrad Wojtasik, Wroclaw University of Science and Technology. Chunking Methods on Retrieval-Augmented Generation: Effectiveness Evaluation Against Computational Cost and Limitations. arXiv 2606.00881, 30 May 2026.
• Olivier Martinez, Sciences Po. Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization, 2023 to 2026. arXiv 2607.14035, 15 July 2026.
• GeoCited. Which Content Formats Win the Most AI Citations for B2B SaaS? 40 prompts on ChatGPT and Google AI Overviews, 490 cited URLs, 28 August 2026, one run per prompt. Our own research, published with its data.