Princeton GEO study: what it found and what to do about it
Researchers at Princeton found that adding statistics, direct quotations, and cited sources to a web page raised its visibility in generative AI search engines by up to 40% in controlled tests. Fluent writing without specifics had no measurable effect. Keyword stuffing performed no better than the unchanged baseline. Aggarwal, N. et al. "GEO: Generative Engine Optimization." KDD 2024; arXiv 2311.09735. arxiv.org/abs/2311.09735. DIRECTIONAL caveat: the exact percentage varied by method and benchmark. Re-verify per-method numbers at the source before citing a specific decimal.
What did the Princeton GEO study actually test?
The researchers took a corpus of web pages and responses generated by AI search engines (including Bing, Perplexity, and You.com), then applied nine targeted interventions to the source pages. They measured whether each intervention changed how often the modified pages appeared in AI-generated answers. The interventions included: adding statistics, adding direct quotations from authoritative sources, citing sources explicitly, rewriting for fluency, adding keyword-dense passages, and simplifying language.
The finding was consistent across engines: adding verifiable specifics (statistics, quotations, citations) was the one class of intervention that measurably raised inclusion rates. Fluency rewrites and keyword additions did not. This matches the architecture of retrieval-augmented generation: an AI engine is selecting passages to include in an answer, and it selects passages it can verify and attribute, not passages that sound good.
What does "up to 40% lift" mean in practice?
The study measured "visibility" as the rate at which a modified page appeared in AI-generated answers compared to the unmodified baseline. The largest single-method gains (statistics + citations together) reached ~40% relative improvement in inclusion rate on some benchmarks. The range was wide across methods and engines. DIRECTIONAL: treat "up to 40%" as a directional ceiling for the highest-performing combination, not a guaranteed outcome for any individual page. Re-verify the exact table values at arxiv.org/abs/2311.09735 for a primary-source citation.
Which interventions had the most effect?
Ranked by effect size in the paper (DIRECTIONAL -- verify exact order at source):
- Statistics + cited sources: highest single intervention effect. A sentence like "77.9% of the top 1,000 sites have no AI crawler block (AuditLamp, 2026-07-05)" beats "most sites do not block AI crawlers."
- Direct quotations from authoritative sources: quoting a named expert or official document raised inclusion rates measurably. The AI engine can verify the attribution.
- Citing sources explicitly: even without a full quotation, naming the source (author, institution, date, URL) improved results.
- Fluency rewrites: minimal measurable effect. Well-written vague sentences performed no better than the baseline.
- Keyword density: no measurable effect. Keyword stuffing performed at baseline. This is the most actionable negative finding: SEO keyword tactics do not transfer to generative-engine visibility.
How does this connect to what AuditLamp checks?
AuditLamp's amp.citable_facts and core.answer_block checks are grounded in this study's findings. The engine flags pages that lack concrete statistics, have no cited source near factual claims, or open with non-specific prose. Run the audit to see which of your pages fail these checks. The fix is not more words -- it is one specific fact, dated, sourced, placed at the top of the relevant page. See how few business sites have pages that pass this test.
Cite this summary
AuditLamp. "Princeton GEO study: what it found and what to do about it." 2026-09-03. auditlamp.com/learn/notes/princeton-geo-study-what-it-means. Primary source: Aggarwal et al., "GEO: Generative Engine Optimization," KDD 2024, arXiv 2311.09735. Free to use with attribution; verify exact figures at the primary source.