Key takeaways
- The paper, "GEO: Generative Engine Optimization", was posted to arXiv in November 2023 and accepted to KDD 2024.
- It built GEO-bench, 10,000 queries from nine sources, and tested nine ways of rewriting a web page on an AI answer engine built for the study and on Perplexity.
- Adding quotations, adding statistics and citing sources worked best. Keyword stuffing scored below the unedited pages on the main metric.
- Gains were biggest for pages ranked fifth in the search results; top-ranked pages often lost visibility. The authors list clear limits.
This summary of the generative engine optimization paper sticks to what the authors wrote. Every number below comes from the arXiv abstract or the full text, version 3. For what GEO is as a practice, and how I apply it, see the GEO hub.
What is the generative engine optimization paper?
It is a research paper by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. First posted to arXiv on November 16, 2023, and accepted to KDD 2024, it proposed generative engine optimization as a way for content creators to raise their visibility in AI-generated answers.
The starting problem is simple. AI answer engines read several web pages and write one answer, so the owner of a page has little control over whether, or how, the page shows up. The abstract claims GEO can boost visibility by up to 40% and that the best strategies change by domain. The current version (v3) is dated June 28, 2024.
How did the GEO study work?
For each query, the researchers fetched the top five Google results and gave them to gpt-3.5-turbo, which wrote an answer citing them. They then rewrote one source page with a GEO method, generated the answer again and compared how visible that page was before and after. Five answers were sampled for each query.
Two metrics measured visibility. Position-Adjusted Word Count counts how many words of the answer come from sentences citing the page, giving more weight to citations that appear early. Subjective Impression averages seven judgments scored by GPT-3.5, such as relevance, influence, uniqueness and the likelihood that a user would click. The rewriting itself was done by a language model following each method's instructions.
What is GEO-bench?
GEO-bench is the benchmark the authors released with the paper: 10,000 queries split into 8,000 for training, 1,000 for validation and 1,000 for testing. The queries come from nine sources, including MS MARCO, ORCAS-1, Natural Questions, ELI5, Perplexity.ai Discover and queries written with GPT-4, across 25 domains.
About 80% of the queries are informational, 10% transactional and 10% navigational, according to the paper. Each query also carries a category tag, assigned with GPT-4 and checked by hand on the test set. Keep that mix in mind: most queries ask for information, not for a business to hire.
Which GEO methods worked best?
Quotation Addition, Statistics Addition and Cite Sources led, with relative gains of 30% to 40% on Position-Adjusted Word Count and 15% to 30% on Subjective Impression. Fluency Optimization and Easy-to-Understand rewrites also helped. Keyword Stuffing scored below the unedited baseline on the main metric, and the authors say such methods offer little to no improvement.
| Method (as named in the paper) | Position-Adjusted Word Count | Subjective Impression |
|---|---|---|
| Quotation Addition | 27.2 | 24.7 |
| Statistics Addition | 25.2 | 23.7 |
| Fluency Optimization | 24.7 | 21.9 |
| Cite Sources | 24.6 | 21.9 |
| Technical Terms | 22.7 | 21.4 |
| Easy-to-Understand | 22.0 | 20.5 |
| Authoritative | 21.3 | 22.9 |
| Unique Words | 20.5 | 20.4 |
| No optimization (baseline) | 19.3 | 19.3 |
| Keyword Stuffing | 17.7 | 20.2 |
Scores copied from Table 1 of the main text, GEO-bench test set. Higher is better. Be aware that the appendix (Table 6) prints a different set of values, with a baseline of 19.8, so summaries that mix the two tables give slightly different percentages.
Do lower-ranked pages benefit more?
Yes, in this study. Citing sources raised visibility by 115.1% for pages ranked fifth in the Google results but lowered it by 30.3% for pages ranked first. Quotation Addition (+99.7% at rank five, -22.9% at rank one) and Statistics Addition (+97.9%, -20.6%) followed the same pattern, so smaller sites gained most.
The authors read this as GEO giving smaller content creators a fairer chance against big sites. The paper also tested domains separately: Statistics Addition did best in Law and Government, Debate and Opinion queries, Quotation Addition in People and Society, Explanation and History, and Cite Sources in Statement, Facts and Law and Government. On a 200-query subset, the best pair of methods, Fluency Optimization plus Statistics Addition, beat any single method by more than 5.5%.
Did the methods work on a real AI engine?
The authors repeated the test on Perplexity.ai with a 200-query sample. Quotation Addition gave the best Position-Adjusted Word Count, a 22% improvement over the baseline, and Statistics Addition raised Subjective Impression by 37%. Keyword Stuffing again performed about 10% worse than the baseline, which confirms its result on the study's own engine.
What are the limits of the GEO study?
The authors list four: only two generative engines were tested, so methods may need to change as engines evolve; queries change over time, so the benchmark needs updates; effects on search ranking were not measured; and the category tags may be noisy. Add one more: the engine and the judge were 2023 models.
That last point is my own reading, not the authors'. The test engine used gpt-3.5-turbo and the top five Google results, and GPT-3.5 also scored the Subjective Impression. Today's AI Overviews, AI Mode and ChatGPT search use other models and fetch far more pages. The study measures how a page is used inside one answer, not whether it is cited at all, nor whether anyone clicks or buys.
What a business should take from the paper
- Add evidence, not adjectives. Statistics with a source and quotes from named experts were the strongest methods.
- Write clearly. Fluency and easy-to-understand rewrites helped; they also help human readers.
- Do not stuff keywords. It scored below the unedited pages on the main metric and on Perplexity.
- Smaller sites have a real opening. The biggest gains went to pages ranked fifth, not first.
- Test on your own questions. Results varied by domain, so measure your buyers' questions instead of trusting a single percentage.
To turn these lessons into writing, read how to write content for AI search, and use the GEO checklist for the technical side. My GEO audit runs your buyers' questions in six AI assistants to see where you stand today, and the home page explains how I work.
Frequently asked questions
Who wrote the GEO paper?
Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. It was accepted to KDD 2024. You can read the full paper on arXiv and see how each method was measured, with which queries and which engines.
Where can I read the GEO paper?
It is free on arXiv under the number 2311.09735, as a PDF and as an HTML page. Version 3, from June 28, 2024, is the current one. Read the methods section to understand how the results were measured. It also lists the limits.
Is the 40% figure guaranteed?
No. The abstract says up to 40%, measured on average across the benchmark with a 2023 test engine. Your pages may gain more, less or nothing. Use it as a guide, not a promise, and measure your own results with real questions.
Is GEO-bench public?
Yes. The authors released it with the paper as a benchmark of 10,000 queries so others could test their own methods. It is useful for research, but your own customer questions are a better test for your business, because they show what really affects you.
Did the paper test Google AI Overviews?
No. It tested its own engine built on gpt-3.5-turbo with Google results, and Perplexity.ai. Google's AI features were not measured. That is why its results cannot be applied directly to Google, and it pays to run your own tests. Use your customers questions.
Want to see these guides first? Choose JJ SEO Colombia as a preferred source on Google, and my articles are more likely to show up for you in Top Stories.
Follow on GoogleAdd as a preferred source