The Princeton GEO Study: Benchmark & Findings Explained

The paper "GEO: Generative Engine Optimization" (Aggarwal, Dugan, et al., KDD 2024) is the first systematic, benchmarked study of how content modifications affect citation visibility in AI-generated answers. It introduced GEO-bench — a benchmark of 10,000 real search queries across 9 datasets — and tested 9 optimization strategies with quantified visibility results. As of July 31, 2026, it remains the most methodologically rigorous reference study in the GEO field, with directional findings validated by Conductor, GrackerAI, BrightEdge, and Axis Intelligence.

The paper was published at KDD 2024 (ACM SIGKDD Conference on Knowledge Discovery and Data Mining), with the arXiv preprint 2311.09735 first posted in November 2023. The research team spanned Princeton University, IIT Delhi, and Georgia Tech — a cross-continental collaboration. In 2026, the paper continues to be cited on Princeton's research portal as foundational work in AI content optimization.

How GEO-bench measures AI visibility

GEO-bench is constructed from 10,000 real search queries drawn from 9 authentic datasets, including Google Search logs and Perplexity user queries. Each query is run through a generative search engine, and the research team analyzes which sources are cited in each AI answer and how prominently.

The core metric is position-adjusted word count: it counts how many words from a source appear in the AI-generated answer, with words appearing earlier in the answer weighted more heavily. This design distinguishes between "briefly mentioned" and "core reference" — a crucial distinction that simple binary citation metrics miss. The metric captures both whether you are cited and how prominently you are cited.

The 9 strategies: complete results table

The research team applied each content modification to a baseline page and compared visibility before and after on GEO-bench. Below are the complete results:

Source: Aggarwal et al., "GEO: Generative Engine Optimization," arXiv:2311.09735, KDD 2024. Visibility measured by position-adjusted word count on GEO-bench (10,000 queries × 9 datasets).

The most counterintuitive finding: lower rank, higher GEO returns

Perhaps the most strategically significant finding from the Princeton study is that GEO's impact is inversely correlated with traditional search rank. Pages ranked 5th in Google search achieved up to +115% visibility improvement after GEO optimization. Meanwhile, the 1st-ranked page experienced a 30% visibility decline.

This finding inverts the traditional SEO logic where top-ranking pages compound their advantage through authority signals. In AI search, the re-ranking models do not inherit Google's ranking conclusions — they evaluate content quality independently. BrightEdge's February 2026 analysis confirms this mechanism at scale: 62–83% of sources cited by Google AI Overviews come from outside the organic top 10. The implication is clear — pages that struggle in traditional SEO have the most to gain from GEO.

"For two decades, SEO was a winner-take-most game — high-authority domains with deep backlink profiles consumed the majority of traffic. GEO may break this pattern. If content quality is strong enough — original data, expert quotations, clear structure — a new domain can earn more exposure in AI answers than established authority sites."

2026 independent validations of the Princeton findings

Since publication, multiple independent studies have validated and extended the Princeton team's findings. Here is a summary of the key validation studies as of July 31, 2026:

Practical application: a 5-step content checklist from the Princeton study

Based on the Princeton findings and their 2026 validations, here is a prioritized checklist for content optimization:

1. 1.
   Add expert quotations (+41% lift) — Interview domain experts with full attribution. Use blockquote formatting with named sources. GrackerAI 2026 reports that attributed quotes increase AI citation probability by 3.7×, the single largest structured data multiplier.
2. 2.
   Replace vague claims with statistics (+33% lift) — Every assertion should include a verifiable number with source and date. Authoritas 2026 found that pages with original statistical data achieve +156% higher citation probability in Google AI Overviews.
3. 3.
   Optimize readability (+29% lift) — Remove redundancy, ensure clear subject-verb structure, limit sentences to 25 words. The fluency + statistics combination outperforms individual strategies by +5.5% (Princeton paper).
4. 4.
   Cite authoritative sources (+28% lift) — Link every factual claim to academic papers, official statistics, or recognized industry reports. University of Toronto 2025 research (arXiv:2509.08919) found that brands cited on third-party platforms receive 6.5× more AI citations than those appearing only on owned domains.
5. 5.
   Stop keyword stuffing (−8% penalty) — AI engines use semantic embedding matching, not keyword counting. BrightEdge 2026 data confirms natural writing outperforms keyword-optimized content for AI citation. Use synonyms and related entities instead.

Study limitations

While the Princeton study is the methodological gold standard in GEO, several limitations merit attention:

• ▸ The experiment used a single generative engine (an open-source GPT-3.5 system), not production ChatGPT Search or Perplexity. Different AI engines may exhibit different citation behaviors.
• ▸ The 9 strategies were tested individually, not in combination (except the fluency + statistics pair). In practice, multiple strategies interact, and the combined effects may differ from individual lifts.
• ▸ 10,000 queries, while substantial, may not represent highly niche or industry-specific search intents. Results may vary by vertical.
• ▸ The benchmark predates the ChatGPT Search launch (October 2024), Google AI Mode (March 2025), and the 2026 shifts in market dynamics. Subsequent validations address these gaps but do not replicate the controlled experimental design.

The directional conclusions should be treated as reliable guidance rather than precise prediction formulas. The convergence of independent validations — Conductor, GrackerAI, BrightEdge, Authoritas, Axis Intelligence — provides confidence in the overall framework while acknowledging that exact lift percentages vary by platform, vertical, and content type.

Frequently asked questions