The Princeton paper, decoded. The academic foundation of GEO.

The paper that named the field

In November 2023, six researchers from Princeton, Georgia Tech, the Allen Institute for AI, and IIT Delhi posted a paper to arXiv titled simply “GEO: Generative Engine Optimization.” A year later it was published at ACM SIGKDD 2024, the top data-mining conference in the world.

The authors — Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande — did three things no one had done before:

1. Coined the term “Generative Engine Optimization” and formalized generative engines as systems combining retrieval with grounded LLM synthesis.
2. Built GEO-bench — a benchmark of 10,000 queries from 9 diverse sources across 25 domains, designed to evaluate optimization methods systematically.
3. Tested nine concrete content optimization strategies and quantified what worked, what didn’t, and by how much.

The paper is the academic spine of the entire GEO discipline. Every credible practitioner study published since either cites it, replicates it, or extends it.

“Including citations, quotations from relevant sources, and statistics can significantly boost source visibility, with an increase of over 40%. — Aggarwal et al., KDD 2024”

The benchmark that made it measurable

Before GEO-bench, “optimizing for AI search” was advice. After GEO-bench, it was a measurable discipline.

The benchmark draws queries from nine sources representative of real-world search behavior:

• MS Marco, ORCAS-1, Natural Questions (commercial & informational queries)
• AllSouls (Oxford academic exam questions)
• LIMA, Davinci-Debate (LLM-generated diverse prompts)
• Perplexity Discover (real production AI search queries)
• ELI-5 (Reddit explanation requests)
• GPT-4 generated queries (synthetic coverage)

The query distribution preserves real-world ratios: 80% informational, 10% transactional, 10% navigational. Coverage spans 25 domains: tech, health, finance, history, law, government, opinion, and more. The split is 8,000 train, 1,000 validation, 1,000 test.

Two evaluation metrics measure visibility: Position-Adjusted Word Count (how prominently a source is quoted, weighted by position in the AI response) and Subjective Impression (LLM-judged influence on the answer).

What worked

The paper tested nine optimization methods. The top three each delivered 30–40% relative improvement on Position-Adjusted Word Count and 15–30% on Subjective Impression:

1. Cite Sources — adding inline references to authoritative third-party sources directly within the content.
2. Quotation Addition — embedding pull quotes from credible experts and publications.
3. Statistics Addition — injecting specific, attributed numerical data into the text.

Stylistic improvements (Fluency Optimization — better readability, clearer structure) added 15–30% on top. The single highest-performing combination was Fluency Optimization + Statistics Addition together, which outperformed any single method by more than 5.5%.

Real-world validation on Perplexity.ai (a deployed production engine, not just the benchmark): Quotation Addition improved Position-Adjusted Word Count by 22%, Statistics Addition improved Subjective Impression by 37%.

What didn’t

Two negative findings deserve as much attention as the positive ones, because they upend SEO instincts:

• Keyword stuffing — the workhorse of early SEO — showed “little to no improvement” on generative engines. Generative engines synthesize meaning, not match terms.
• Authoritative / persuasive tone changes — “writing more confidently” — showed no significant lift. The authors note GEs are “already somewhat robust to such changes,” meaning models discount confidence-as-style and weight grounded substance.

The pattern: tactics that game surface-level patterns don’t work. Tactics that add genuine substance (citations, real statistics, structured quotation, clear writing) work measurably and reproducibly.

The asymmetry that changed everything

The single most consequential finding for the GEO industry is buried in the experimental section: lower-ranked sites benefit dramatically more from GEO than top-ranked sites do.

• SERP rank-5 sites applying Cite Sources gained +115% visibility increase.
• Top-ranked sites on average LOST 30.3% visibility when applying the same methods.

The interpretation: when everyone plays the same game, the playing field levels. Smaller creators — previously crushed by domain authority moats — gain disproportionately. The historic SEO advantage of sheer authority compresses inside generative engines because synthesis pulls from a wider source pool than ranked lists.

This is why Aggarwal et al. wrote: “The advent of Generative Engines might initially seem disadvantageous to these smaller entities. However, the application of GEO methods presents an opportunity for these content creators to significantly improve their visibility.” The paper is, in a real sense, an invitation to the underdogs.

The follow-ups: Toronto and SourceBench

Two academic papers in 2025–2026 extended the foundation in important directions.

Chen et al., University of Toronto (2025) — “Generative Engine Optimization: How to Dominate AI Search” — ran the first comprehensive comparative analysis of AI Search vs Google Search across multiple verticals, languages, and query paraphrases. The headline finding: AI Search exhibits “systematic and overwhelming bias toward earned media” — in US automotive queries, AI search returned 81.9% earned media vs 18.1% brand-owned and 0% social, compared to Google’s much more balanced 39.5% brand / 15.4% social / 45.1% earned mix. This validated, at scale, what Visigent’s Signal Hierarchy methodology codified separately: third-party citations are the highest-leverage GEO investment.

SourceBench (2026) evaluated source quality across 12 AI search systems. Two findings stand out: GPT-5 achieves the highest source quality scores, and AI search actively discovers high-quality sources not found in traditional keyword-based search results. The implication: GEO isn’t just optimization for the same set of sources Google ranks. It’s competition for a partially different source pool.

Why the paper still matters

Three years after first publication, every reputable GEO claim still routes back to GEO-bench. The 40% visibility lift number is now a category benchmark. The +115% lift for lower-ranked sites is the underlying math behind why the discipline works for challenger brands. The Fluency + Statistics combination is the operational recipe most credible practitioners reference, often without realizing they’re paraphrasing Aggarwal.

If you take one thing from the paper: add specific, attributed statistics to your content. Of the nine tactics tested, this single move plus fluent writing produced the largest measurable gain. The discipline starts here.

Every statistic in this piece follows Visigent's research standards: named sources, dated figures, corrections on request to help@visigent.com.