Princeton GEO paper (Aggarwal et al., KDD '24) — the foundational generative engine optimization study

Claim: Aggarwal et al. (Princeton + IIT Delhi + Georgia Tech + Allen AI) introduced "Generative Engine Optimization" as a discipline in arXiv:2311.09735 (v1: Nov 2023; v3: Jun 2024), accepted at KDD '24 Barcelona (ACM DOI 10.1145/3637528.3671900). They proposed GEO-bench (10,000 queries × 9 source datasets × 25 domains) and tested 9 optimization methods on GPT-3.5-turbo + a 200-query Perplexity.ai validation subset.

Source: arXiv:2311.09735 v3 — https://arxiv.org/abs/2311.09735. KDD proceedings paper.

Confidence: Verified (primary).

Why it matters for Candid: This is the most rigorous publicly available study on what content patterns lift AI-response visibility. The headline finding (+30-40% lifts) is from this paper. Every "GEO" claim in industry blog posts traces back here.

Atomic findings filed separately: GEO finding: Quotation Addition is the top-performing tactic at +41% on Position-Adjusted Word Count, GEO finding: Statistics Addition is the #2 tactic at +31%, GEO finding: Cite Sources lifts visibility +28%, GEO finding: Keyword stuffing is the only tested tactic that hurts AI visibility (-8% to -10%), GEO finding: Lower-ranked pages (organic rank ~5) gain 115.1% AI visibility when given GEO treatments.

Caveats: Tested on 2024-era engines (GPT-3.5, Perplexity). Whether the lifts persist on Gemini 3.5 / GPT-5 / Claude 4 / Sonar Pro is unverified. The Subjective Impression metric has been critiqued (Sandbox SEO) for construction.