How Do You Need to Structure Your Content So an AI Cites It?

With three changes to form, not content: consistent heading hierarchy, information split into lists and tables instead of long paragraphs, and the direct answer at the start of each section instead of the end. A University of Tokyo study measured that reorganizing just that — without changing a single word of the content — raises citation rate by up to 17.3% across six AI engines.

Of the 39 Argentine e-commerce sites Nostos analyzed in its sector study, 62% score below 30 out of 100 in content structure. It's the worst-performing criterion in the whole sector — worse than authority, worse than technical markup, worse than data density. And it's the only one of Nostos's eight criteria that doesn't depend on what you say, but on how you arrange it.

What Nostos measures when it talks about "structure"

The Tokyo paper proposes a framework (GEO-SFE) that splits content structure into three levels, and Nostos evaluates all three:

• Macro: document-level architecture — heading hierarchy (H1, H2, H3, with no skipped levels or disorder) and how evenly organized the sections are.
• Meso: organization within each section — whether information sits in run-on paragraphs or is split into lists, tables and chunks a model can process piece by piece.
• Micro: sentence-level features — where emphasis appears, where the keyword sits, whether the answer is at the start of the sentence or buried at the end.

No AI engine "reads" a site the way a person does. A language model processes the document looking for these three levels of organization before deciding what to cite.

Form matters independently of content

What makes this study methodologically solid is that it isolated the variable: it kept the semantic content intact — same data, same facts — and only changed how it was organized. The paper itself frames this as a formal constraint on its optimization: that the structural transformation preserve the semantic integrity of the original content.

The result, measured across six commercial generative engines: 17.3% more citations and an 18.5% improvement in perceived quality of the responses that incorporated that content.

Why most sites fail at this

In Nostos's study of 39 leading Argentine e-commerce domains, content structure was the sector's weakest criterion: 62% of the sites analyzed scored below 30 out of 100. It isn't a problem of how good the information on those sites is — many have correct, up-to-date data. It's that the information lives in long paragraphs, with no clear heading hierarchy, no lists or tables a model could chunk and cite piece by piece.

What you can do today

Translated into practice, the Tokyo study's three levels come down to this:

• Consistent heading hierarchy: one H1 per page, H2 for each main section, no jumping from H1 to H4.
• Replace long paragraphs with lists, tables and short blocks wherever the content is enumerable — not for aesthetics, but so a model can cite the exact part it needs without having to parse an entire paragraph.
• Put the direct answer at the start of each section, not the end — the model prioritizes what appears first.