Tested nine content-optimization methods on GEO-bench (10,000 queries) and reported visibility gains of up to about 40% in generative-engine responses.
Key: Relative gains in a visibility metric on the authors' benchmark (top-5 Google results fed to gpt-3.5-turbo), not citation rates. Keyword stuffing did not perform well.
Quotation addition +41% (position-adjusted word count)
Automatic rule extraction from generative engines via LLMs + reinforcement learning. Average improvement of about 36% over the strongest Princeton method. Introduces GEU (utility) metric.
How we use it: Informs scoring weight updates. Answer-first structure and passage density checks are based on AutoGEO findings.
How we use it: Schema richness check differentiates rich vs generic. Generic gets zero bonus.
AI Citations Report 2026
OtterlyAI — 1M+ citations analyzed
73% of sites have technical barriers blocking AI crawlers. Reference-grade content gets 3-5x more citations.
How we use it: robots.txt, CDN, and JS rendering checks address the 73% barrier problem.
AI Mode Citation Factors
SE Ranking, 2025 — 2.3M pages
Flesch-Kincaid Grade 6-8 is the sweet spot. Sections of 100-150 words are optimal. FAQ schema has zero impact — only real FAQ content matters.
How we use it: Readability, passage density, FAQ-in-content, and section length checks.
GEO Optimizer focuses on infrastructure optimization (robots.txt, llms.txt, schema, meta, content structure) — not content manipulation. The adversarial research confirms that manipulative techniques degrade the ecosystem.