The GEO paper explained
“GEO: Generative Engine Optimization” (arXiv 2311.09735) is the November 2023 paper that named GEO. Researchers then at Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi built a 10,000-query benchmark (GEO-bench), tested nine content-rewriting methods in a simulated engine built on gpt-3.5-turbo, and reported visibility gains of up to 40%, with citations, quotations and statistics working best and keyword stuffing not working.
Who wrote the GEO paper and where was it published?
Authors: Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande.[1] The v1 PDF (November 16, 2023) gives affiliations as Princeton University, Georgia Tech, The Allen Institute for AI and IIT Delhi.[2] The KDD ’24 version lists Rajpurohit and Kalyan as independent.[1] DOI: 10.1145/3637528.3671900.[3] Code and the benchmark are public.[4]
How did the experiment work?
- Benchmark: GEO-bench, 10,000 queries from nine sources (MS MARCO, ORCAS-I, Natural Questions, AllSouls, LIMA, Davinci-Debate, Perplexity.ai Discover, ELI5 and GPT-4-generated queries), split 8,000/1,000/1,000, about 80% informational.[1]
- Engine: top five Google results per query, answered by gpt-3.5-turbo, five samples at temperature 0.7.[1]
- Intervention: one randomly chosen source per query rewritten by an LLM using one of nine methods.[1]
- Metrics: Position-Adjusted Word Count (share of answer words attributed to a source, discounted by position) and a Subjective Impression score judged by GPT-3.5 using G-Eval.[1]
What were the nine GEO methods?
Authoritative, Statistics Addition, Keyword Stuffing, Cite Sources, Quotation Addition, Easy-to-Understand, Fluency Optimization, Unique Words and Technical Terms.[1]
| Method | Position-adjusted word count (baseline 19.3) | Subjective impression avg (baseline 19.3) |
|---|---|---|
| Quotation Addition | 27.2 | 24.7 |
| Statistics Addition | 25.2 | 23.7 |
| Fluency Optimization | 24.7 | 21.9 |
| Cite Sources | 24.6 | 21.9 |
| Keyword Stuffing | 17.7 | 20.2 |
Selected rows from Table 1 of the KDD version.[1]
What are the paper’s main findings?
- Best methods improved on baseline “by 41% and 28% on Position-Adjusted Word Count and Subjective Impression respectively.”[1]
- Keyword stuffing performed worse than baseline on word count.[1]
- A persuasive, authoritative tone gave “no significant improvement.”[1]
- Lower-ranked sources gained most; the top-ranked source lost visibility under several methods.[1]
- Fluency Optimization plus Statistics Addition was the best pair, beating any single method by more than 5.5% (on a 200-example subset).[1]
What are the limitations?
The authors note methods “may need to adapt over time as GEs evolve” and that they “didn’t evaluate how GEO methods affect search rankings.”[1] Further limits visible in the design: a simulated engine on one older model; the metric measures how much of the answer draws on a source, not whether users see or click it; the Perplexity test used uploaded files rather than live web retrieval.[1] Later work tested the methods with citation rank on newer models and found most ineffective; see GEO replications.
Frequently asked questions
What is GEO-bench?
A public benchmark of 10,000 queries from nine datasets across 25 domains, released with the GEO paper, each paired with cleaned text of top Google results.
Did the GEO paper test real AI search engines?
Mainly a simulated engine (Google top-5 results plus gpt-3.5-turbo). It also ran a 200-sample test on Perplexity.ai using sources supplied as file uploads.
Is the 40% figure a traffic increase?
No. It is a relative gain in a word-count-based visibility metric inside generated answers, not traffic or clicks.
See also
References
Sources accessed October 14, 2026 unless another date is given. See all sources.
- “GEO: Generative Engine Optimization”. arXiv (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande). v1 submitted November 16, 2023; published at KDD ’24.
- “GEO: Generative Engine Optimization (v1 PDF, author affiliations)”. arXiv. v1, November 16, 2023.
- “GEO: Generative Engine Optimization”. Proceedings of the 30th ACM SIGKDD Conference (KDD ’24). August 2024.
- “GEO-optim/GEO (code and GEO-bench)”. GitHub. accessed October 14, 2026.