Generative engine optimization (GEO): what the research actually says.

The paper, GEO: Generative Engine Optimization by Pranjal Aggarwal and colleagues, was first posted in November 2023 and published at KDD 2024. It named “generative engines”, search tools that write an answer from several sources, and proposed GEO as a way for content creators to improve how visible their content is in those answers.

The idea travels under several names. Wikipedia's generative engine optimization article lists answer engine optimization (AEO), LLM optimization and AI SEO among them, and the article averaged about 8,300 page views a month from September 2025 to August 2026, according to Wikimedia's pageview data.

The researchers built their own generative engine: for each query, it took the top five Google results and had GPT-3.5 Turbo write an answer with citations. They assembled a benchmark, GEO-bench, of 10,000 queries drawn from nine sources, including real search queries, Reddit's ELI5 questions and Perplexity's Discover feed, across 25 domains. Then, for each query, they picked one of the five sources at random and had a language model rewrite it using one of nine methods:

Visibility was scored two ways: a position-adjusted word count (how much of the answer drew on the source, weighted toward the top), and a subjective impression score that GPT-3.5 rated on factors such as relevance, influence and how likely a reader would be to click.

• Substance won. The authors report that the best methods improved on the baseline by 41% on position-adjusted word count and 28% on subjective impression, with citations, quotations and statistics leading.
• Readability helped too. Fluency and easy-to-understand rewrites also lifted visibility, which suggests presentation matters as well as content.
• Keyword stuffing did not. It scored below the baseline on the word-count measure; the authors found it offered little to no improvement.
• The gains went to lower-ranked sources. Adding citations raised visibility by 115.1% for sources ranked fifth in the search results, but lowered it by 30.3% for sources ranked first.
• What works depends on the question. Citations helped most on factual questions and law and government; quotations on people and society, explanations and history; statistics on law and government, debates and opinions.

The test on a real engine

The authors also ran some methods on Perplexity.ai. Quotation addition improved the word-count measure by 22%, other methods showed gains of up to 9% and 37% on the two measures, and keyword stuffing performed 10% worse than the baseline.

This is the part most summaries leave out. The up-to-40% figure is a result on the authors' own benchmark, under their setup. It is not a forecast for any website, including yours.

• A lab engine, mostly. The main results come from a research engine built on five Google results and GPT-3.5, not from Google AI Overviews or ChatGPT. The one live test, on Perplexity, showed smaller gains for some methods.
• Machine rewrites and machine judges. A language model rewrote the content, and GPT-3.5 scored the subjective measure.
• Visibility, not business results. The study measured presence in answers, not clicks, inquiries or revenue.
• No ranking test. The authors say they did not evaluate how the methods affect search rankings.
• A moving target. The authors expect methods to need updating as generative engines change.
• A relative game. Gains were measured against other sources in the same answer, and top-ranked sources sometimes lost ground.

Read as a whole, the study points toward content that is better sourced, more specific and easier to read. Those are good changes for human readers too, which is the safest kind of optimization.

• Cite your sources, and link to the primary document where you can.
• Use real statistics, each with a named source and a date. Never invent one to sound authoritative.
• Quote people and documents that matter, briefly and exactly, with attribution.
• Write clearly. Plain, fluent language helped in the study and helps every reader.
• Drop keyword repetition. It was the one method that made things worse.
• Match the method to the page: citations for factual pages, quotations for people and explanations, data for opinion and debate.
• Make sure the engines can reach the page, because none of this matters if a crawler is blocked.
• Test it on your own questions, with a dated record of what each engine cited before and after the change.

Content like this also tends to compound over time; our guide to content marketing strategies that compound covers the broader program.

GEO builds on SEO rather than replacing it. The engines in the study pulled their sources from search results, so a page still has to be found first. What changes is the standard for the page itself: the study suggests that being quotable, sourced and specific matters more than repeating the words in the query. Our AI search optimization page explains how we approach both halves.