A March 2026 preprint from the University of Tokyo and University of Tsukuba tested structural variants of the same content across six generative engines. The words, claims, and facts were kept identical. Only the structure changed. The result: a mean 17.3% improvement in citation rates across all six engines.
Evidence tier: Preprint finding — directional, not yet peer-reviewed. The 17.3% figure is a mean across six generative engines. Individual engine performance varies. Treat this as an orientation indicator, not a guaranteed outcome.
The paper — "Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior" by Junwei Yu, Yang MuFeng, Yepeng Ding, and Hiroyuki Sato (arXiv: 2603.29979) — is the first controlled experiment to isolate structural variables as an independent lever on AI citation behaviour. The practical implication is direct: you can test structural changes on your existing pages without rewriting a word of content.
What the earlier research measured — and what this paper adds
The original GEO research from Aggarwal et al. 2024 — the peer-reviewed study behind the Evidence Density Score — measured what happens when you combine multiple content modifications at once: adding statistics, adding source citations, improving readability, and changing structure together. That study found associations between the combined bundle of changes and higher citation probability. What it could not isolate was the structural contribution on its own.
The Yu et al. preprint addresses that directly. By holding semantic content constant — same words, same claims, same facts — and varying only structural treatment, the study isolates the structural contribution. The 17.3% figure is what structural changes alone produced when the content itself was unchanged. Our read: this is the cleanest isolation of structure as an independent variable the GEO literature has produced so far.
The three-level hierarchy
The paper organises structural features into three levels. Understanding the distinction matters because the levels operate at different scales and suggest different types of changes to your pages.
Macro-structure is document architecture — how the page as a whole is laid out at the top level. This includes whether a document has a clear introduction and conclusion, how major sections are sequenced, and whether the overall information flow is organised from general to specific or vice versa. First impression. Macro-structure is what a generative engine parses when it first retrieves a page.
Meso-structure is information chunking — how content within sections is grouped and divided. One idea per paragraph: that is the operational rule at this level. Paragraph boundaries, the internal coherence of each block, and the logical grouping of related claims all operate here. A page where each paragraph carries a single complete idea has different meso-structure from a page where ideas run across multiple paragraphs without clear breaks.
Micro-structure is visual emphasis — the local markers that indicate what matters within a passage. Headings, bold text, and lists all operate at this level. These are the properties most visible to a reader scanning a page, and the ones that AI systems can parse most directly as structural markers.
The controlled experiment tested variants of pages across all three levels simultaneously. The 17.3% mean citation improvement represents the combined effect of structural treatment at all three levels — the study does not break out which level contributed most. That platform-level and level-specific breakdown is a gap in the available data from this preprint.
What this means for your pages
The finding is actionable without a content rewrite. You can audit the structural treatment of an existing page — how it is laid out, how paragraphs are chunked, how headings and emphasis are distributed — and make changes that leave your voice and content intact. If the Yu et al. finding holds at peer review, structural changes alone can move citation rates in a measurable direction.
For independent niche operators, this matters because rewriting content is the highest-cost GEO intervention. Changing what you know and how you express it takes time and risks losing the specificity that makes niche content worth citing in the first place. Structural changes — reorganising sections, tightening paragraph boundaries, adding headings to break up large blocks of text — are lower-cost interventions that the Yu et al. study suggests have measurable citation lift on their own. The qualification is that the preprint evidence is directional. This is the right direction to test, not a guaranteed result.
What this study does not tell us
The study is a preprint from March 2026. It has not completed peer review, which means the methodology and conclusions have not been independently validated. The 17.3% figure is a mean across six engines — the study summaries available do not report individual engine performance, so the variance across platforms is not visible in the current data. Perplexity, ChatGPT, and Google AI Overviews are known to behave differently in citation selection (see Zhang et al. 2026), and the structural gains may not be evenly distributed across them. The paper also does not establish whether structural gains are durable across model updates — a question the field has not yet answered for any GEO intervention. Treat this finding as a directional preprint result that warrants acting on cautiously while peer review proceeds. We are doing exactly that with the tools it validates.
The Psytable tools this finding validates
Two Psytable tools surface the structural properties the Yu et al. research identifies as citation levers. The Heading Visualiser maps the heading structure of any page, showing how document architecture and micro-structure work at the macro and local levels simultaneously. The AI Snippet Extractor identifies which passages in a page are most likely to be extracted and cited by a generative engine — a direct indicator of how chunking and emphasis affect citation selection. We built both tools around the structural dimensions the Yu et al. preprint places at the centre of citation behaviour.
References
Yu, Junwei, Yang MuFeng, Yepeng Ding, and Hiroyuki Sato (2026), "Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior," preprint. arXiv: 2603.29979
Aggarwal, Manas et al. (2024), "GEO: Generative Engine Optimization," ACM SIGKDD 2024. arXiv: 2311.09735
Audit your page structure.
The Heading Visualiser maps your document architecture and micro-structure — the two structural levels the Yu et al. preprint identifies as citation levers. No signup. No gates.