Generative engine optimization (GEO) is a term introduced in a 2023 paper by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, which proposed it as the first framework for helping content creators improve their visibility in the responses of generative engines.[1] First posted in November 2023, the paper was accepted to the KDD 2024 conference; its first version listed affiliations with Princeton University, Georgia Tech, the Allen Institute for AI and IIT Delhi.[1][2]
Generative engines
The paper defines generative engines as systems that answer queries by retrieving information from multiple sources and summarizing it with large language models, and models them as a set of generative models combined with a search engine that returns sources.[3] It argues that such engines leave creators little control over whether and how their content appears, because the systems are opaque and change quickly.[1]
GEO-bench
The authors built GEO-bench, a benchmark of 10,000 queries drawn from nine sources including MS MARCO, Natural Questions, ELI5 and Perplexity.ai's Discover feed. About 80% of the queries are informational, and each is paired with the text of the top five Google results.[3]
Visibility metrics
The paper measured visibility in two ways. Position-adjusted word count counts the words in sentences citing a source, weighted toward sentences that appear earlier in the response. Subjective impression combines seven qualities scored by a language model, including relevance, influence, uniqueness and likelihood of a click.[3]
Methods and results
The study tested nine ways of rewriting source content, using an answer engine built on GPT-3.5 and a subset of queries on Perplexity.ai.[3] Results on position-adjusted word count, compared with an unmodified baseline score of 19.5:[3]
| Method | Score | Effect |
|---|---|---|
| Quotation addition | 27.8 | Best single method |
| Statistics addition | 25.9 | Improved |
| Cite sources | 24.9 | Improved |
| Keyword stuffing | 17.8 | Worse than baseline |
The authors reported that the best methods improved visibility by about 41% on position-adjusted word count and 28% on subjective impression, and that combining fluency optimization with added statistics worked best of all.[3] Lower-ranked sources gained far more than top-ranked ones: citing sources raised the visibility of the fifth-ranked result by 115% while reducing the first-ranked result's visibility by 30%.[3] Effectiveness varied by subject; added statistics helped law, government and opinion questions, while quotations helped questions about people, society and history.[3]
Limitations
The authors noted that they tested only two generative engines, that effective methods may change as engines evolve, that the benchmark will need updating, and that they did not measure how GEO edits affect conventional search rankings.[3]
Later research
Subsequent studies have tested and questioned GEO's findings. A 2025 benchmark, C-SEO Bench, found most conversational SEO methods largely ineffective and often harmful, while traditional SEO that raised a source's rank in the material the model reads worked much better, and found that gains shrank as more sites adopted the same methods.[4] A 2025 preprint reported that AI search tools strongly favor earned, third-party media over brand-owned content.[5]
Other researchers have examined manipulation. A 2024 EMNLP paper showed that prompt-injection text on web pages could promote low-ranked products in conversational search, including on Perplexity,[6] and a 2024 preprint described "preference manipulation attacks" against Bing and Perplexity, arguing that widespread use would degrade results for everyone.[7]
References
- ^ ^ ^ ^ ^ ^ Aggarwal, Pranjal; Murahari, Vishvak; Rajpurohit, Tanmay; Kalyan, Ashwin; Narasimhan, Karthik; Deshpande, Ameet (2023). "GEO: Generative Engine Optimization" (arXiv:2311.09735). arXiv (accepted to KDD 2024). arxiv.org
- ^ Aggarwal, Pranjal; et al. (November 16, 2023). "GEO: Generative Engine Optimization (version 1)" (v1). arXiv. arxiv.org
- ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ Aggarwal, Pranjal; et al. (June 28, 2024). "GEO: Generative Engine Optimization (full text)" (v3). arXiv. arxiv.org
- ^ Puerto, Haritz; et al. (2025). "C-SEO Bench: Does Conversational SEO Work?". arXiv (NeurIPS 2025 Datasets and Benchmarks). arxiv.org
- ^ Chen, Mahe; et al. (2025). "Generative Engine Optimization: How to Dominate AI Search". arXiv (preprint). arxiv.org
- ^ Pfrommer, Samuel; Bai, Yatong; Gautam, Tanmay; Sojoudi, Somayeh (2024). "Ranking Manipulation for Conversational Search Engines". arXiv (EMNLP 2024). arxiv.org
- ^ Nestaas, Fredrik; Debenedetti, Edoardo; Tramèr, Florian (2024). "Adversarial Search Engine Optimization for Large Language Models". arXiv (preprint). arxiv.org