The short version
- In a study of 200 articles across six AI engines, Yu and colleagues (opens in a new tab) reported that restructuring alone lifted the citation rate from 45.0% to 52.8%.
- In 252,000 head-to-head trials across six AI models, Vishwakarma and colleagues (opens in a new tab) found formatting choices had no impact on which source was cited first.
- In our study of pages cited by AI Overviews, pages with a table were no more likely to be cited, but once cited were 14.7 points more likely to have their wording used.
- In an analysis of 602 prompts by Zhang and colleagues (opens in a new tab), pages in Q&A format had slightly lower influence on answers, a difference of -5.74%.
Can changing a page’s structure alone raise AI citations?
In one lab study, yes, by about a sixth. Yu and colleagues (opens in a new tab) took 200 articles from a standard research dataset and rewrote their structure while keeping the meaning. They then tested original and restructured versions across six AI engines, including ChatGPT, Claude and Perplexity.
The share of queries citing the content rose from 45.0% to 52.8%. In the authors’ words, the six engines showed consistent 17.3% citation improvements. Gains ranged from 14% to 20% across different kinds of engine.
The articles were long, averaging 2,547 words, and came paired with 377 real-world queries. An automated check confirmed the restructured versions kept their meaning. The paper is a preprint, and it does not explain in detail how it queried the live engines. Two engine names it lists, Google SGE and Bing Chat, are older product names.
Which kinds of structure mattered most in that study?
The big-picture layout mattered most, and visual emphasis least. The authors split structure into three levels and removed each one in turn to see how much the gain fell.
| Level of structure | What it covers | Share of the gain |
|---|---|---|
| Document architecture | Heading hierarchy, navigation and overall flow | 44.9% |
| Information chunking | Paragraphs, lists and tables | 39.7% |
| Visual emphasis | Bold, italics and keyword placement | 15.4% |
An AI judge also scored the restructured pages higher on “influence” (+32.0%) and “click probability” (+31.4%). These are ratings by an AI model, not clicks measured from real users.
Do other controlled tests find the same thing?
No: the largest controlled test found formatting made no difference. Vishwakarma and colleagues (opens in a new tab), at the software company Sprinklr, ran 252,000 trials across six AI models. Each trial showed a model two product pages that differed in exactly one factor.
They report that formatting choices had no impact, suggesting the AI models parse content regardless of visual organization. Dense paragraphs versus organized sections, and grouped versus scattered information, did not change which page was cited first.
What did matter were facts and fit: matching the topic, stating a price, carrying a recent date and appearing first in the list. Those four won across all six models. We cover them in our guide to why AI picks a competitor’s page. Note this test used only two pages at a time with brand names removed, unlike a live search.
A 2025 benchmark by Puerto and colleagues (opens in a new tab) points the same way. A combined rewrite that included bolding and restructuring gained only in one domain, retail products, on one model.
Does structure help a page get chosen, or help the AI use it?
The best evidence says it helps use more than choice. In our study of pages cited by AI Overviews, we compared 3,096 top-10 pages on the same searches. Tables, headings and other layout features showed no reliable link to being cited.
But structure mattered after the choice was made. Cited pages with an HTML table were 14.7 points more likely to have their citing passages absorbed, meaning the AI Overview’s wording could be traced to the page. Tables showed no association with being cited at all (+0.6). Tables plausibly hold the specific figures an answer repeats.
Zhang and colleagues (opens in a new tab) found a similar pattern in a public dataset of 602 prompts across ChatGPT, Google and Perplexity. Pages that most shaped answers averaged 10.59 headings, against 0.85 for pages that shaped them least. They caution this is a snapshot that cannot show whether headings cause the effect. Other traits of the pages engines lean on are described in our guide to content AI engines favor.
Do FAQ sections and question-style headings help?
Not reliably. In the Zhang dataset, pages in Q&A format had a mean influence of 0.0947 against 0.1005 for other pages. The authors conclude the value comes from the evidence inside the page, not from question marks in headings.
Our study found the same. Question-form headings, compared with pages on the same search, showed a gain of +3.5 points that could not be told apart from zero. FAQ headings showed no gain either. We look at this more closely in our guide to FAQ pages and AI citations.
Does structured data count as structure that helps?
The evidence is mixed and mostly against a lift. A 2025 study by Kumar and Palkhouski (opens in a new tab) audited 1,100 pages cited for 70 B2B software prompts. It found metadata, clean page markup and structured data most associated with citation. Its authors list a private research company alongside a university. Every page it audited had been cited somewhere, so it compared cited pages with each other, not with pages passed over.
When we compared cited pages with uncited pages on the same search, Organization schema showed no meaningful association with citation (+0.5 points). Only a machine-readable date held up, at +7.9 points, and mainly on informational searches. In a before-and-after test that our study reports, Ahrefs tracked 1,885 pages that added schema against 4,000 controls. Adding schema did not raise citations. The wider set of page features is reviewed in our guide to on-page signals and citations.
What should you do about it?
Restructure to make facts easy to find and reuse, but do not expect structure alone to get you chosen.
- Fix substance and fit first. Topic match, prices, specifications and dates beat formatting in the largest controlled test.
- Use clear headings and logical sections. Document layout carried the most weight in the one study that found a structural lift, and it helps human readers too.
- Put key figures in tables. In our data, tables did not help a page get cited, but they went with more of the answer coming from the page.
- Don’t convert everything to FAQs. Q&A packaging alone showed no gain in two separate studies.
- Add a machine-readable date. It was the one page feature that held up in our comparison, mainly for informational content.
- Keep ranking in view. In our study, position in Google’s results mattered more than any page feature.
If you want help prioritizing these edits across your site, see our generative engine optimization service.
What does the research not tell us yet?
No one has yet restructured real pages and measured citations on live AI search before and after.
- The positive result is one preprint. The 17.3% lift comes from a single study whose method for querying live engines is not described in detail.
- The negative result is vendor research. The 252,000-trial study comes from a software company and used two anonymized pages at a time.
- The observational studies cannot show cause. Our study and the Zhang analysis compare pages as they are; well-structured pages may also be better in other ways.
- “Use” is measured by word overlap. Our absorption measure tracks wording, not whether the answer is faithful to the page.
- No study measured clicks or sales. None of these studies tracked whether restructured pages brought more visitors or customers.
Frequently asked questions
Do headings and bullet points help content get cited by ChatGPT?
The evidence is split. One lab study found a 17.3% citation lift from restructuring, while a 252,000-trial test found formatting had no impact.
Should I add tables to my pages for AI search?
Yes, where you have real figures to show. In our study, tables did not raise the chance of citation, but cited pages with tables were 14.7 points more likely to have their wording used.
Is FAQ formatting good for AI visibility?
Not on its own. Q&A pages had slightly lower influence on AI answers in one dataset, and FAQ headings showed no gain in our comparison.
Is restructuring cheaper than writing new content for AI search?
It can be, but the payoff is uncertain. The only study showing a structural lift was a lab test, and no live before-and-after test exists yet.
Sources
- Yu, MuFeng, Ding and Sato (2026), Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior (opens in a new tab), arXiv:2603.29979.
- Vishwakarma, Kumar and Jamidar (2026), What Gets Cited: Competitive GEO in AI Answer Engines (opens in a new tab), arXiv:2605.25517.
- Zhang, He and Yao (2026), From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms (opens in a new tab), arXiv:2604.25707.
- Puerto, Gubri, Green, Oh and Yun (2025), C-SEO Bench: Does Conversational SEO Work? (opens in a new tab), arXiv:2506.11097.
- Kumar and Palkhouski (2025), AI Answer Engine Citation Behavior An Empirical Analysis of the GEO16 Framework (opens in a new tab), arXiv:2509.10762.
- Underneath (2026), What pages cited by AI Overviews have in common