Page structure helps AI find your content. A little, and not the way you think.

Document structure helps an AI system retrieve the right passage, but only modestly, and mostly for one reason: where the text is cut. That is the result of a study published on 7 October 2026 by three researchers at the University of Notre Dame. Its most telling test borrows a tool from medicine. They gave a retrieval system a placebo.

The placebo was a set of headings that looked real and belonged to other documents. Same format, same depth, same length, wrong content. If headings helped simply by adding words to a passage, the fake ones would have helped too. They did not. On one of the two collections, they made retrieval worse.

The question matters for anyone who publishes content AI systems read. Answer engines and retrieval-augmented generation (RAG) systems do not read a page whole. They split it into passages, called chunks, turn each into a vector, and retrieve the closest ones to a question. How a page is cut and labelled changes what can be found.

What did the Notre Dame study test?

The study, arXiv:2610.10170, compares four document-structure techniques used in RAG systems under a single protocol, on two very different collections. Its authors are Andrey Kuehlkamp, Priscila Correa Saboia Moreira and Samuel Rund of the Center for Research Computing at the University of Notre Dame. The arXiv listing describes it as an initial draft: it has not been through peer review.

The four techniques are the ones vendors and developers argue about most. The first cuts documents at their section boundaries instead of into fixed-size windows. The second asks a language model to write a short context sentence for each chunk, the Contextual Retrieval method Anthropic published in 2024. The third prefixes each chunk with its heading path: document title, then section, then subsection. The fourth searches in two stages, choosing sections first and passages second.

The collections were chosen to be far apart. One holds 200 Wikipedia Featured Articles with 951 generated questions, filtered so they could not simply copy the wording of the answer. The other holds 1,585 scientific papers from the QASPER dataset, with 4,303 questions written by real readers about 1,508 of them. Eleven conditions ran on both, and four comparisons were fixed before the results came in.

Two controls make the study unusual. Fixed-size windows were matched, document by document, to the average length of the structural chunks, within four tokens, so that a difference in cutting could not hide a difference in size. And the placebo arm prefixed each chunk with a real heading path taken from another document, so the effect of the heading's content could be separated from the effect of simply adding tokens.

Does cutting a document at its headings help retrieval?

Yes, and it is the largest effect in the study: re-cutting the same text at section boundaries, with nothing added, raised the retrieval score from 0.362 to 0.403 on Wikipedia and from 0.098 to 0.117 on the scientific papers. The score is coverage-nDCG@10, which measures how much of the evidence for a question appears in the first ten passages, giving more credit to passages ranked higher.

Adding the true heading path on top of those structural chunks added much less: 0.011 on Wikipedia, 0.009 on the papers. On both collections, the change of cut was worth two to four times the heading prefix. The same ordering held with a smaller embedding model, so it is not an artefact of one encoder.

Structural cutting also found more of the evidence, not just in a better order. On the scientific papers, the share of evidence passages present among the top 50 results rose from 0.248 with fixed windows to 0.302 with structural chunks and real heading paths.

Do headings help because of their words, or because of what they say?

Because of what they say: true heading paths beat the placebo on both collections, by 0.010 on Wikipedia and 0.016 on the scientific papers. The placebo headings matched the real ones in format, depth and length. Only their content was wrong.

The comparison with plain structural chunks, carrying no heading at all, is even clearer. On Wikipedia, the placebo bought exactly nothing: both scored 0.403. On the scientific papers, it hurt, falling from 0.117 to 0.109. A heading that does not describe its section is at best neutral and at worst noise. The authors write that the prefix's value depends on its being true of the chunk.

Does an AI-written summary on each passage help retrieval?

Less than cutting at the structure, and it costs space. Fixed windows with a generated context sentence scored 0.391 on Wikipedia and 0.114 on the papers. Plain structural chunks, with no added text at all, scored 0.403 and 0.117. The generated context did not recover what badly placed boundaries had lost.

The cost shows when every token counts. In a 512-token context budget, where the added sentences take room that evidence would otherwise fill, the contextualized windows fell below plain fixed windows on both collections: 0.167 against 0.187 on Wikipedia, 0.046 against 0.052 on the papers. The generated contexts were capped at 80 tokens. The authors also note that the gains Anthropic reported for its method came from an internal benchmark that was not peer-reviewed.

Does searching section by section work better than a single search?

No: choosing sections first and passages second lost to a single search over the same passages, by 0.033 on Wikipedia and 0.015 on the scientific papers. Both systems used identical vectors, so the loss comes entirely from the routing step.

The authors traced it to the first stage. The sections it kept contained the right evidence in only 60% to 68% of cases on Wikipedia, and 17% to 20% on the papers, where it had to choose among roughly 22,000 sections. Whatever the first stage drops, the second cannot recover. A reranking model narrowed the gap on the papers from 0.0147 to 0.0038. The authors limit the conclusion to this common design, not to hierarchical retrieval in general.

Does the author's own structure beat a structure rebuilt by an AI?

On Wikipedia yes, on the scientific papers no. The researchers also stripped every heading and asked a 26-billion-parameter open model, Gemma-4-26B, to rebuild the structure. On Wikipedia, the authors' headings beat the rebuilt ones by 0.0129, a significant gap. On the papers, the difference was 0.0011, statistically nothing.

The model did not simply copy the original sections. Only 78.9% of its chunks on Wikipedia, and 75.8% on the papers, matched an original chunk exactly. The authors offer a hypothesis, untested: depth. 61% of the papers have a single level of sections, while every Wikipedia article in the set had at least three. A flat structure leaves little for the author's version to add.

How strong are the results?

The effects are small but consistent: three of the four comparisons hold with the same sign and statistical significance on both collections. The table reports them as published.

In standardized terms, the gains are between 0.06 and 0.11, which statisticians call small. The authors are direct about it: they claim reliability, not magnitude. They also list the limits. Both collections are in English. A single model rebuilt the structures. The Wikipedia questions were generated by a model. And only retrieval was measured: whether these gains change the quality of the final answer is outside the study.

What does this change for a brand's web pages?

It argues for pages built from real sections with headings that describe them, not for more headings. What follows is the KAIS reading of the study. It extends a result on RAG systems to web pages by analogy, and says so where it does.

One section should cover one subject. The study's largest effect came from where the text is cut. If a system splits a page at its headings, a section that mixes fees, method and team becomes one muddled passage, close to no question in particular. A heading should mark a real change of subject, not decorate a long block.

A heading should tell the truth about its section. The placebo result is the clearest lesson here. Headings such as "Our approach", "Why choose us" or "Excellence" say little about the text below them. By inference, they sit closer to the placebo than to a true heading path. "Fees for registering a company in Thailand" or "How long a work permit takes" describe what follows. Whether question-form headings do better than descriptive ones is a separate question this study did not test.

The structure should be explicit in the code. On the deep Wikipedia articles, the authors' headings beat a structure rebuilt by a model. Real heading tags (h2, h3) spare a system the guesswork, and long pages with several levels have the most to gain. Nesting deeper is not the goal: in the authors' exploratory analysis, most of the benefit came from the first heading level and then plateaued.

Expectations should stay modest. These are gains of one or two hundredths on a retrieval score. Structure helps a passage be found. It does not make a source trusted, and it does not guarantee that a cited source shapes the answer, a gap measured in Cited is not used.

What if you build your own AI assistant?

For a company running its own RAG assistant on its documents, the authors give a direct order of priorities. Cut where the author cut. Prefix each chunk with one level of real heading path. Spend the language-model budget on rebuilding structure where it is missing rather than on writing a context for every chunk. And do not route questions to sections first unless a reranking model checks the results behind it.

That order fits the setting the study measured: long, structured documents and strong open embedding models. A firm that answers client questions from its own guides, contracts or policies is close to that setting.

Where does this sit in the KAIS AI Visibility Chain?

The study sits at the second and third stages of the KAIS AI Visibility Chain, Be Retrieved and Be Understood: it measures whether the right passage can be found, and whether its label describes it.

It says nothing about the stages that follow. A passage that is easy to retrieve can still come from a source an engine does not trust, be cited without shaping the answer, or never lead to a recommendation. The last of those stages, and when it happens in a conversation, is the subject of The AI waits before it recommends you.

A heading is a promise about the text below it

The study does not show that structure is a shortcut to AI visibility. It shows something narrower and more useful. Retrieval improves when a document is cut where its subjects change, and when each part carries a label that is true.

The placebo headings had the right format, the right depth and the right length. They failed for the only reason that counted: they were not true of the text beneath them.

Sources

• Kuehlkamp, Moreira and Rund — Does Document Structure Help Dense Retrieval? A Placebo-Controlled Ablation of Four Mechanisms Across Two Corpora, arXiv:2610.10170 [cs.IR], 7 October 2026, Center for Research Computing, University of Notre Dame. Initial draft, not peer-reviewed. Pre-registered contrasts (Table 1), per-condition scores (Table 2), token-budget results (section 6.2), placebo analysis (section 6.3), section-recall diagnostic (section 6.4), gold vs induced structure (section 6.5), exploratory depth analysis (section 6.6), practical recommendations (section 7) and limitations (section 8). All figures in this article come from the paper; none were reproduced or re-measured by KAIS.
• Anthropic — Introducing Contextual Retrieval, 2024. The chunk-contextualization method tested as one arm of the study, cited by its authors.