Guide · 10 min read
The GEO research paper: what Aggarwal et al. tested and found
Almost every article about generative engine optimization cites one source: the GEO research paper by Aggarwal and colleagues, first posted in November 2023 and published at KDD 2024. Most cite a single number from it. This guide goes through the whole paper: who wrote it, how the benchmark was built, the nine methods it tested, the figures it actually reports, the limits the authors admit, and what a later benchmark found when it tested the same ideas.

What the GEO research paper is
The paper is titled GEO: Generative Engine Optimization. Version 1 appeared on arXiv on 16 November 2023; version 3, accepted at KDD 2024 in Barcelona, on 28 June 2024. The paper and its code are public.
The GEO research paper did two things. It named and formalised a problem: AI answers summarise several sources, and site owners have little control over whether theirs is one of them. And it measured whether simple rewrites change how much of an answer is built from a page.
Our guide to how to write content AI cites summarises it in a few lines. The details change how much weight each finding deserves.
Who wrote it, and why it is called the Princeton study
The six authors are Pranjal Aggarwal and Vishvak Murahari (listed as equal contributors), Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. The affiliations differ between versions, which explains some confusion online:
- Version 1 (November 2023) lists Princeton University for Murahari, Narasimhan and Deshpande, Georgia Tech for Rajpurohit, the Allen Institute for AI for Kalyan, and IIT Delhi for Aggarwal.
- Version 3, the KDD 2024 version, lists IIT Delhi for Aggarwal, Princeton University for Murahari, Narasimhan and Deshpande, and 'Independent, Seattle' for Rajpurohit and Kalyan.
Half the authors are at Princeton in both versions, hence the shorthand Princeton GEO study.
How the experiment worked
The authors built a simplified answer engine modelled on tools like Perplexity. For each question it took the top five Google results, passed the full text of those five pages to GPT-3.5-turbo, and asked it to write an answer with a citation after every sentence. Each answer was sampled five times to reduce noise.
For each question, one of the five sources was picked at random to be rewritten. It was rewritten once with each method, the answer regenerated, and its visibility compared with before. GPT-3.5-turbo did the rewriting, using prompts published in the authors' repository.
That design matters. Retrieval, the step where a real assistant decides which pages to read, was held fixed. The experiment measures how a page competes once it is in the model's context, not how it gets there.
GEO-bench: the 10,000 questions
To test the methods, the authors released GEO-bench, a benchmark of 10,000 queries split into 8,000 for training, 1,000 for validation and 1,000 for testing. The results in the paper are on the test split. The queries come from nine sources:
- MS MARCO, ORCAS and Natural Questions, which contain real anonymised queries from Bing and Google.
- AllSouls, essay questions from All Souls College, Oxford.
- LIMA (reasoning questions) and Davinci-Debate (debate questions).
- Trending queries from Perplexity's Discover section and questions from the ELI5 subreddit.
- Queries generated by GPT-4 to add variety.
The authors say the benchmark keeps a real-world mix of 80% informational queries and 10% each of transactional and navigational ones, and covers 25 domains such as arts, health and games. GPT-4 tagged every query along seven dimensions such as difficulty and intent; the appendix warns these tags can be noisy.
How the GEO research paper measures visibility
In a list of links, visibility is mostly rank. In a written answer, a source can be cited in passing or carry half the paragraph. The GEO research paper therefore defines its own impression metrics:
- Word Count: the share of the answer's words that sit in sentences citing the source. A sentence cited by several sources is split between them.
- Position-Adjusted Word Count: the same, with more weight for sentences near the top of the answer, using an exponential decay. This is the paper's main objective metric.
- Subjective Impression: seven scores produced by GPT-3.5 in the G-Eval style, covering relevance, influence, uniqueness, diversity, the chance a user follows up or clicks, perceived position and perceived amount of content. The scores are rescaled to match the first metric.
All scores are normalised so the five sources in an answer add up to 100. With five sources, an average one sits near 20. Keep this in mind when you read the tables: none of these metrics counts clicks, sales or whether a brand was recommended.
The nine methods tested
Each method is a type of rewrite that an LLM applied to the source text:
- Authoritative: a more persuasive and authoritative style.
- Statistics Addition: quantitative statistics instead of qualitative statements where possible.
- Keyword Stuffing: more keywords from the query, as in classic SEO.
- Cite Sources: relevant citations from credible sources.
- Quotation Addition: quotations from credible sources.
- Easy-to-Understand: simpler language.
- Fluency Optimization: more fluent text.
- Unique Words: unusual words where possible.
- Technical Terms: technical vocabulary where possible.
Most only change presentation; keyword stuffing, citations and quotations can require adding material.
The results, with the paper's own figures
On the test split, the unmodified sources scored 19.3 on Position-Adjusted Word Count. The figures below are the paper's Table 1:
- Quotation Addition: 27.2, the best result. The caption says the best methods improve on the baseline by 41% on this metric and 28% on Subjective Impression.
- Statistics Addition: 25.2.
- Fluency Optimization: 24.7. Cite Sources: 24.6.
- Technical Terms: 22.7. Easy-to-Understand: 22.0. Authoritative: 21.3.
- Unique Words: 20.5.
- Keyword Stuffing: 17.7, below the unmodified text.
In the text: citations, quotations and statistics gave 30 to 40% relative improvement on the main metric and 15 to 30% on the subjective one, and fluency and readability rewrites gave 15 to 30%. A more authoritative tone brought no significant improvement in the authors' reading, and keyword stuffing offered little to no gain. The abstract rounds the headline to 'up to 40%'.
Two more analyses matter. Combining methods helped: Fluency Optimization plus Statistics Addition was the best pair, beating any single method by more than 5.5%, measured on a subset of 200 examples. And the gains depended on where the page started. Cite Sources raised the visibility of the fifth-ranked result by 115.1% while the top result fell by 30.3%; Quotation Addition and Statistics Addition showed the same pattern, with gains of 99.7% and 97.9% for the fifth result.
Domain differences
Table 3 lists the query categories where each method worked best:
- Authoritative: debate, history and science.
- Fluency Optimization: business, science and health.
- Cite Sources: statements, facts, and law and government.
- Quotation Addition: people and society, explanations and history.
- Statistics Addition: law and government, debate and opinion.
The authors recommend domain-specific adjustments. The categories come from automatic tags and no per-category sample size is given, so treat this as a direction to test.
The test on Perplexity
The paper also ran the methods on Perplexity.ai with 200 test questions. Because Perplexity does not accept chosen URLs, the authors uploaded the source texts as files. It tests how Perplexity writes from given documents, not how it finds pages.
Quotation Addition again led on the main metric, 29.1 against a baseline of 24.1, which the authors describe as a 22% improvement. Statistics Addition reached 33.9 on Subjective Impression against 24.7, the 37% figure. Keyword stuffing scored 21.9 on the main metric, which the authors describe as about 10% worse. The appendix tables also show results the summary leaves out: Cite Sources scored 19.0 on Subjective Impression, below the 24.7 baseline, and keyword stuffing scored 28.1 on that metric, above it.
Limitations, the authors' and ours
The limitations section of the GEO research paper names four: answer engines will change; real queries change over time; the effect on search rankings was not tested; and the query tags may contain errors. We would add:
- The engine was GPT-3.5-turbo reading five pages. Current assistants use different models, read more sources and decide for themselves what to retrieve.
- Only one source was optimised per question. Nobody else in the answer was competing with the same tactics.
- Visibility is the share of words, judged partly by another LLM. It is not a recommendation, a click or a sale.
- The rewrites were written by an LLM. Statistics and quotations added this way have to be real and sourced on a live site, or they become a liability.
What later research found
The most direct follow-up we could verify is C-SEO Bench: Does Conversational SEO Work? by Puerto and colleagues, accepted at NeurIPS Datasets and Benchmarks 2025. It tested methods from the GEO research paper, including Authoritative, Statistics, Citations and Fluency, across question answering and product recommendation, in six domains, and with several competitors optimising at once.
Its abstract reports that most of these methods were largely ineffective and frequently lowered a document's ranking in the answer. Improving a document's position in the model's context, the job of classic SEO, mattered more than any rewrite, and the gains shrank as more competitors adopted the same methods. The authors attribute part of the gap with the original paper to the metric: word share versus citation ranking.
The paper also cites Kumar and Lakkaraju (2024), who showed that a crafted text sequence on a product page could push it to the top of an LLM's recommendations. The GEO authors call their own methods non-adversarial.
What practitioners should and should not take from it
The GEO research paper supports a few modest conclusions:
- Specific, sourced facts help. Real statistics, named sources and accurate quotations were the most consistent winners, and they also make a page more useful to people.
- Clear writing helps. Fluency and readability rewrites improved visibility without new content.
- Keyword repetition does not help, and on the main metric it hurt.
- Being retrieved still comes first. The paper only tested pages that were already among the five sources, and the later benchmark found that position in context dominates.
What it does not support: a promise of 40% more AI visibility, claims about ChatGPT, Gemini or AI Overviews (not tested), or the idea that rewrites replace SEO. If someone sells a fixed uplift based on it, ask which metric and which engine.
The practical route is to apply the robust findings and then measure on your own questions. Start with how AI search works, choose the questions with prompt research, and track the results as described in how to measure AI visibility. For where the term GEO sits next to AEO, see AEO vs GEO and our introduction to what GEO is.
Frequently asked questions
Is the GEO research paper from Princeton?
Partly. Three of the six authors are at Princeton University in both versions. Version 1 also lists Georgia Tech, the Allen Institute for AI and IIT Delhi; the KDD 2024 version lists IIT Delhi and two independent researchers in Seattle.
Where does the '40% more visibility' figure come from?
From the abstract, which says GEO can boost visibility by up to 40%. The best single result, Quotation Addition, improved the paper's Position-Adjusted Word Count metric by 41% over unmodified text on GPT-3.5-turbo. It is a best case on one metric, not an average or a guarantee.
Did the study test ChatGPT or Google AI Overviews?
No. The main experiments used GPT-3.5-turbo reading the top five Google results, plus a smaller test on Perplexity.ai with uploaded files. Results on other assistants were not measured.
Is keyword stuffing bad for AI search?
In this study it lowered visibility on the main metric, both on the GPT-3.5-turbo setup and on Perplexity. On Perplexity it scored higher on the subjective metric, so the evidence is not clean, but no result in the paper makes a case for adding keywords.
Can I use GEO-bench to test my own pages?
The benchmark, code and prompts are public. For a business, it is usually more useful to track your customers' real questions in the assistants they use, month after month.


