Every figure is attributed, linked, and answer-ready — engineered to be cited.
New to this space? Start with the complete field guide to AI search optimization.
Scale and trajectory — from primary sources and named industry research. These are the figures most frequently cited in GEO discussions.
Five Ahrefs large-sample studies on what signals predict AI citation, how overlap with traditional rankings has changed, and what content formats AI engines prefer.
The most rigorous controlled research on GEO comes from Aggarwal et al., published at ACM KDD 2024 — the first peer-reviewed paper to isolate the effect of specific GEO tactics on AI citation rates.
The KDD 2024 paper is the closest thing AI search optimization has to a randomized controlled trial: it isolated individual content tactics (adding statistics, adding quotations, rewriting for fluency, restructuring answer-first) and measured AI citation lift independently. The +115% / −30% figures capture the range of effect — they are not averages across all pages. For full context, read the manifesto on what the data actually shows.
AI citations are volatile in ways that SEO rankings are not. Two data points that reframe how you should measure GEO progress.
These two figures together explain why share-of-answers across many repeated runs is the only honest metric in GEO. A rank tracker shows you a single position. An AI answer check on any given day can be right roughly 8 times out of 9 — or wrong on that day precisely. The GEO vs SEO comparison covers what this means for measurement in depth.
Two data points from a live GEO program — framed as measured, not caused. Confounders exist; the numbers are real.
External stats (Sections A–B) are sourced from: (1) primary corporate disclosures — Alphabet earnings calls, SEC filings; (2) Gartner press releases; (3) Ahrefs large-sample studies with disclosed methodology (sample sizes and date ranges are included in each citation). No figure is taken from vendor marketing copy, survey extrapolation without primary disclosure, or aggregate round-ups without traceable origins.
Peer-reviewed research (Section C): Aggarwal et al. (2024), "GEO: Generative Engine Optimization," ACM SIGKDD. The paper is available on arXiv. The ±% figures reported are within-study comparisons of GEO-optimized vs. control pages at different initial rank positions; they should not be extrapolated to all pages or all AI engines.
First-party data (Sections D–E): LumenGEO and GrantCompass data are my own. They are explicitly labeled First-party and framed as measured (what was observed) not caused (what produced the outcome). These are operator observations, not controlled proofs.
What's excluded: Any stat of the form "X% of AI citations come from [source type]" that cannot be traced to a named study with disclosed methodology. Vendor-produced statistics without independent verification. Survey-based predictions without primary links.
The full pipeline — engines, signals, retrieval gates, and a working glossary of GEO terminology.
Start here →87 experiments, 135,700 citations, and what the numbers actually show — including what doesn't move citations.
Read the findings →Same foundation, different unit of competition. What transfers, what breaks, and how to allocate effort.
Read the comparison →Score any URL on the 14 signals that predict AI citation — with ranked fixes, in 30 seconds.
Scan your page →Run any URL through the free GEO Readiness Scanner — it checks 14 structural signals against the patterns this research identifies, with the exact fixes ranked by impact. Or talk to LumenGEO about owning this for your site.
Updated June 2026 · Built to be cited: sourced, structured, and answer-first.