RESEARCH PAPER — SIGI-2026-037 — March 2026

Structural Correlates of AI Citation: An Observational Analysis of 22 On-Page Metrics

Q: Are these correlations evidence of causation?

No. Gate 2 of the Logic-First Methodology (Confound Check) fails completely for this dataset. Every variable that differs between cited and uncited groups differs simultaneously. The cited sites are older, more established, have more content, and more backlinks. No individual correlation can be isolated as causal.

V. Tavitian, J. Tavitian — The Scientific Institute for Generative Intelligence

Category C — Competitive Intelligence • Evidence Level 3 • Observational Correlation

This paper presents observational correlations between 22 on-page metrics and AI citation outcomes across 21 anonymised service websites. H2 count shows the strongest positive correlation (cited average 14, uncited average 6, ratio 2.3x). H2 question percentage shows the strongest negative correlation (cited 6%, uncited 54%, ratio 0.1x). Schema type count is paradoxically higher in uncited sites. All correlations are maximally confounded by site age, domain authority, and training data presence. Gate 2 (Confound Check) of the Logic-First Methodology fails completely. No causal claims are supported by this analysis.

Which Metrics Correlate Positively with AI Citation?

Ten of the 22 metrics show higher averages in cited sites compared to uncited sites. H2 count demonstrates the strongest positive structural correlation at 2.3x (cited: 14, uncited: 6). Statistics count shows a 2.0x ratio (cited: 4, uncited: 2). Paragraph count shows 1.6x (cited: 44, uncited: 28). Social proof word count shows 1.6x (cited: 14, uncited: 9). Word count shows 1.4x (cited: 2,486, uncited: 1,725).

Image count shows the most extreme ratio at 30.7x (cited: 368, uncited: 12), but this figure is heavily skewed by portfolio-heavy sites in the game outsourcing vertical and should not be interpreted as a general relationship between image count and citation probability.

Metric	Cited Avg (5+)	Uncited Avg (0)	Difference	Ratio	Direction
H2 Count	14	6	+8	2.3x	CITED HIGHER
Statistics Count	4	2	+2	2.0x	CITED HIGHER
Paragraph Count	44	28	+16	1.6x	CITED HIGHER
Social Proof Words	14	9	+5	1.6x	CITED HIGHER
Word Count	2,486	1,725	+761	1.4x	CITED HIGHER
Named Entity Count	16	12	+4	1.3x	CITED HIGHER
H2 Question %	6%	54%	−48	0.1x	UNCITED HIGHER
Schema Type Count	7	11	−4	0.6x	UNCITED HIGHER
FAQ Question Count	1	5	−4	0.2x	UNCITED HIGHER
Hreflang Count	0	4	−4	0.0x	UNCITED HIGHER

Table 1. Selected correlations from the 22-metric analysis. Cited group: sites scoring 5+. Uncited group: sites scoring 0. All correlations confounded.

Which Metrics Show Counter-Intuitive Negative Correlations?

Several metrics show higher averages in uncited sites, producing counter-intuitive findings. Schema type count averages 11 in uncited sites versus 7 in cited sites (0.6x ratio). FAQ question count averages 5 in uncited sites versus 1 in cited sites (0.2x ratio). Hreflang count averages 4 in uncited sites versus 0 in cited sites. Internal link count averages 37 in uncited versus 21 in cited (0.6x ratio).

These inverse correlations likely reflect compensatory optimisation: newer sites that have not yet achieved citation may invest more heavily in technical SEO signals (schema, hreflang, FAQ markup) as an attempt to accelerate visibility. The established cited sites achieved citation through content authority and domain age, not through schema density.

Why Does Gate 2 Fail for This Analysis?

Gate 2 of the Logic-First Methodology requires that observed differences between groups can be attributed to the variable of interest rather than confounding factors. In this dataset, Gate 2 fails completely because every variable that differs between cited and uncited groups differs simultaneously. The cited sites are older, have higher domain authority, have more training data presence, and have been building content and backlinks for years. No individual on-page metric can be isolated as a causal contributor to citation when all metrics co-vary with site maturity.

The only valid interpretation is associational: these metrics are associated with citation outcomes in this sample. Whether any metric causally contributes to citation requires the controlled experimental designs proposed in SIGI-2026-035.

Suggested Citation

Tavitian, V. & Tavitian, J. (2026). Structural Correlates of AI Citation: An Observational Analysis of 22 On-Page Metrics. The Scientific Institute for Generative Intelligence, SIGI-2026-037. https://generativeintelligence.institute/publications/SIGI-2026-037/

Frequently Asked Questions

Which on-page metric shows the strongest positive correlation with AI citation?+

H2 count shows the strongest positive structural correlation: cited sites average 14 H2 headings versus 6 for uncited sites, a 2.3x ratio. However, this correlation is maximally confounded by site age and domain authority.

Does schema markup correlate positively with AI citation?+

Surprisingly, schema type count is higher in uncited sites (average 11) than cited sites (average 7). This counter-intuitive finding may reflect compensatory optimisation by newer sites.

Are these correlations evidence of causation?+

No. Gate 2 (Confound Check) fails completely. Every variable that differs between groups differs simultaneously. No individual correlation can be isolated as causal.

APPLIED RESEARCH

Apply This Research

For applied GEO services based on this research → TDS GEO Agency

TDS GEO Agency Contact SIGI