What 2 million AI citations show: prompt-content match matters most, checklists matter less
Once you control for the domain, the usual AEO checklist (FAQ blocks, structured data, Core Web Vitals) reverses or collapses to zero — domain authority and prompt alignment dominate.
Key numbers
- Sample: roughly 2 million citations from four commercial engines (including ChatGPT, Claude and Gemini) over six months, joined to 10,000 crawled pages across 19 B2B SaaS workspaces, with 60+ page features tested.
- Nine analytical methods were combined for consensus, including mixed-effects regression with domain fixed effects, FDR correction, stability-selection Lasso, double machine learning, generalised additive models and temporal hold-out replication.
- The most robust finding: prompt-content alignment (Jaccard overlap between page tokens and the workspace's full prompt corpus) is the dominant page-level predictor, with beta = +0.37 (95% CI [0.33, 0.41], q ≈ 10^-73).
- The negative result that matters: the standard AEO checklist (FAQ blocks, structured data, Core Web Vitals) shows positive effects in pooled data but reverses or collapses to zero once domain fixed effects are applied — a textbook Simpson's paradox.
- Domain-level AI authority exceeds the strongest non-alignment page-level feature by a factor of six in mean absolute SHAP value.
Method and sample
arXiv paper “What Drives Citations in Production Large Language Models? An Observational Multi-Method Study of Two Million AI Citations Across Ten Thousand Web Pages” (2026-09-28): ~2 million citations from four commercial engines joined to 10,000 crawled pages from 19 B2B SaaS workspaces, with 60+ features tested through nine methods and only fully surviving findings reported. The engine set includes Google AI, which is outside our monitoring scope.
What it means for brands
Two practical consequences: pages have to actually answer the questions customers ask (prompt alignment), which is the strongest page-level variable; and the structured-data checklist is not a master key — it can look effective across sites and then disappear once the domain is controlled for. What does not wash out is domain authority: what third parties say about you.
How SourceSight handles it
SourceSight's prompt library is the customer question corpus, and the content factory rewrites pages against those questions (alignment first); the off-site list handles the domain-authority layer by market and business type. We still produce structured data — we just do not sell it as the only lever.
Limits
The sample concentrates on B2B SaaS workspaces and English prompts, so extrapolation across industries needs care; modelling domain fixed effects deliberately absorbs variance that page-level factors would otherwise claim, which is a methodological trade-off rather than proof that page features do nothing.
References
Run it on your own project
What the same prompts return across 10 engines is only knowable by running them.
Related research
Whose side is the AI shopping assistant on? Change its role and the penalty for sponsored results disappears
The same sponsored listings get treated very differently once the prompt names the platform as the agent's principal — disclosure rules built for humans do not carry over.
How engines work · 2026-10-07Ask “which should I buy” and ChatGPT vs Gemini share 0% of their sources in 76.7% of cases
AI shopping advice is neither stable nor consistent: ask again and the products change, and the API shows something different from the product.