Structured pages get cited more by AI engines because retrieval systems select passages based on extractable structure, not just topical relevance. This is not a theory. Independent research across six generative engines and more than 100,000 citation events, including a March 2026 arXiv preprint, supports it. In Machine Relations, the structural layer that determines whether a source survives retrieval is called citation architecture.
This article maps the findings from five independent research efforts — none affiliated with Machine Relations — to the citation architecture framework. The evidence converges on a single conclusion: structural optimization is now a measurable, predictable driver of AI citation probability.
The core finding: structure produces a 17% citation rate improvement #
The most direct test comes from GEO-SFE, a structural feature engineering framework published in March 2026. Researchers decomposed content structure into three hierarchical levels — macro-structure (document architecture), meso-structure (information chunking), and micro-structure (visual emphasis) — and measured their impact on citation probability across six mainstream generative engines.
The result: a 17.3 percent improvement in citation rate and an 18.5 percent improvement in subjective quality, achieved through structural optimization alone while preserving semantic content (Yu et al., 2026).
This matters because the study held content constant. The improvement came entirely from how information was organized, not from what the information said. That is the citation architecture thesis in quantitative form.
Schema markup is the strongest single content-feature predictor #
A separate large-scale study — "The SEO Floor" — analyzed 100,411 AI citation events from four production platforms (ChatGPT, Perplexity, Claude, Google AI Mode) across 2,000 user queries. The researchers ran per-feature regressions on seven pre-registered GEO levers and found that schema markup had the largest association with citation among the seven features, with an odds ratio of 1.31 per standard deviation, estimated alone beside Google rank tier and vertical (Lee, 2026).
The full feature ranking:
| Feature | Odds ratio |
|---|---|
| Schema markup presence | 1.31 |
| Primary source score | 1.12 |
| Author byline attribution | 1.12 |
| Answer-first coverage | 1.09 |
| Comparison signals | 1.06 |
| List structure | 1.04 |
| Stats density | 1.03 |
| Heading density | 0.94 |
Schema retained an odds ratio of 1.29 in the multivariate model controlling for the other six features and SEO tier. The same study found that a page ranked in Google's top three had 7.82 times the citation odds of a page ranked 11–30, and about 34 times the odds of a page ranked 31–100. Its author calls Google rank the gate. Add schema on pages already inside that gate: the association was estimated beside rank tier and vertical, on 100,411 citation events.
The GEO-16 framework confirms: metadata, semantic HTML, and structured data lead #
An independent auditing framework called GEO-16, published in September 2025, evaluated 1,702 citations across three engines (Brave Summary, Google AI Overviews, Perplexity) using a 16-pillar scoring system. The researchers audited 1,100 unique URLs and found that pillars related to metadata and freshness, semantic HTML, and structured data showed the strongest associations with citation (Kumar et al., 2025).
Pages with a normalized GEO score of 0.70 or higher combined with at least 12 pillar hits aligned with substantially higher citation rates. The framework operationalizes what citation architecture describes: machine-legible structure is not a bonus — it is a threshold requirement.
Why flat-chunk retrieval fails: the structure-preservation evidence #
Most retrieval-augmented generation pipelines linearize HTML documents into fixed-size chunks before indexing. This process destroys section structure, lists, tables, and the contextual scaffolding that makes passages interpretable.
Researchers at SPIRE (Structure-Preserving Interpretable Retrieval of Evidence) built an alternative: a retrieval pipeline that operates over tree-structured documents, preserving structural identity while contextualizing selections within their document hierarchy. Across HTML question-answering benchmarks, this approach yielded higher-quality, more diverse citations under fixed token budgets than strong passage-based baselines (SPIRE, 2026).
The implication for publishers: if you build pages that lose their meaning when stripped of structure, you are building pages that lose their meaning during retrieval. Citation architecture is the editorial discipline that prevents this.
Citation absorption depends on extractable evidence #
The distinction between being cited and being absorbed into an answer is critical. A page that appears in a citation list but contributes nothing to the generated answer has visibility without influence.
Research on citation absorption across 602 controlled prompts, 21,143 citations, and 18,151 successfully fetched pages found that high-influence cited pages — those that actually shape the generated answer — share specific structural properties. They tend to be longer, more structured, semantically aligned with the generated answer, and richer in extractable evidence such as definitions, numerical facts, comparisons, and procedural steps (Zhang et al., 2026).
Citation architecture addresses exactly this: making claims extractable, not just findable.
| Page property | Association in Zhang et al. (mean influence, pages with vs without) |
|---|---|
| Definition markers | +57% |
| Numbers and statistics | +62% |
| Comparison content | +55% |
| How-to content | +41% |
| Q&A format | −6% |
| Semantic alignment with the generated answer | r = 0.36 (answer–citation embedding similarity) |
These associations are among pages already cited, in one snapshot of the 18,151 fetched pages. Add definitions, numbers, comparisons and how-to steps; those were the properties with higher mean influence.
The study also documented platform divergence: Perplexity and Google cite more sources on average, while ChatGPT cites fewer but shows substantially higher average citation influence among fetched pages. In this snapshot, structure and evidence were associated with higher influence among pages already cited. The paper ties selection to authority, recognizability, language and domain context, so put those signals on the page with the structure.
What this means for Machine Relations practitioners #
The convergence across these independent studies validates the citation architecture framework that Machine Relations has operationalized since 2026. The practical translation:
Answer-first structure is measurable. Answer-first coverage was associated with citation at an odds ratio of 1.09 per standard deviation, estimated alone beside Google rank tier and vertical. Put the answer first. That is the citation-architecture move this association supports, on the SEO Floor sample of 100,411 events.
Schema markup is infrastructure, not optimization. Schema presence had the highest odds ratio (1.31 per standard deviation) of the seven features. Treat schema as infrastructure that makes a page machine-addressable. Google's article structured data documentation and the Schema.org Article specification are the starting points.
Macro, meso, and micro structure all matter. The GEO-SFE framework's three-level decomposition maps directly to citation architecture's emphasis on document architecture (H1/H2 hierarchy), information chunking (evidence blocks, tables, lists), and visual emphasis (bold claims, inline definitions). In the paper's ablation, macro-structure contributed 44.9 percent of the citation gain, meso-structure 39.7 percent and micro-structure 15.4 percent, and the authors read contributions that sum to 100 percent as the three levels operating largely independently.
Structure must survive compression. SPIRE's finding that tree-structured retrieval outperforms flat-chunk baselines explains why pages built with clean HTML semantics outperform visually similar pages built with divs and spans. The retrieval pipeline can only preserve structure it can parse.
Entity chains amplify citation architecture. Citation architecture makes individual pages extractable. Entity chains make the brand behind those pages verifiable across independent domains. The combination — structurally citable pages corroborated by cross-domain entity presence — is what separates sources that get cited once from sources that become reusable references.
Where the structural threshold applies #
How concentrated AI citations are depends on what is measured. In the Machine Relations Index's per-category measurement, the top 25 domains took more than 30 percent of a category's citations in only two of 20 reported categories, and reaching 80 percent of a category's citations took more than 400 domains in 16 of the 20 (category concentration study). Most citations sit in a long middle that a publisher can actually enter.
For a domain competing in that middle, citation architecture is the minimum structural standard Machine Relations recommends. The studies above support parts of it, each on its own engines and sample; whether a page meets what a given engine's retrieval favors still has to be measured.
Frequently Asked Questions #
What is the core finding: structure produces a 17% citation rate improvement? #
The most direct test comes from GEO-SFE, a structural feature engineering framework published in March 2026. Researchers decomposed content structure into three hierarchical levels — macro-structure (document architecture), meso-structure (information chunking), and micro-structure (visual emphasis) — and measured their impact on citation probability across six mainstream generative engines.
What is schema markup is the strongest single content-feature predictor? #
A separate large-scale study — "The SEO Floor" — analyzed 100,411 AI citation events from four production platforms (ChatGPT, Perplexity, Claude, Google AI Mode) across 2,000 user queries. The researchers ran per-feature regressions on seven pre-registered GEO levers and found that schema markup was the strongest content-feature predictor of citation, with an odds ratio of 1.31 (Lee, 2026).
What is the GEO-16 framework confirms: metadata, semantic HTML, and structured data lead? #
An independent auditing framework called GEO-16, published in September 2025, evaluated 1,702 citations across three engines (Brave Summary, Google AI Overviews, Perplexity) using a 16-pillar scoring system. The researchers audited 1,100 unique URLs and found that pillars related to metadata and freshness, semantic HTML, and structured data showed the strongest associations with citation (Kumar et al., 2025).
Why flat-chunk retrieval fails: the structure-preservation evidence? #
Most retrieval-augmented generation pipelines linearize HTML documents into fixed-size chunks before indexing. This process destroys section structure, lists, tables, and the contextual scaffolding that makes passages interpretable.
Where does the structural threshold apply? #
How concentrated AI citations are depends on what is measured. In the Machine Relations Index's per-category measurement, the top 25 domains took more than 30 percent of a category's citations in only two of 20 reported categories, so most citations sit in a long middle that a publisher can enter. For a domain competing there, citation architecture is the minimum structural standard Machine Relations recommends.
Sources #
- Yu, J. et al. (2026). "Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior." arXiv:2603.29979.
- Lee, A. (2026). "The SEO Floor: Measuring Google Rank Distribution of AI-Cited Pages." AI+Automation Research.
- Kumar, A. & Palkhouski, L. (2025). "AI Answer Engine Citation Behavior: Bringing the GEO-16 Framework in B2B SaaS." arXiv:2509.10762.
- SPIRE (2026). "Structure-Preserving Interpretable Retrieval of Evidence." arXiv:2604.20849.
- Zhang, K., He, X. & Yao, J. (2026). "From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms." arXiv:2604.25707.