GEO (generative engine optimization) is the work of making your content more likely to be quoted, cited or linked inside answers written by AI search tools such as ChatGPT search, Perplexity, Microsoft Copilot and Google’s AI Overviews. The term comes from a research paper, and it is worth knowing both what the paper measured and what the big platforms say about it.
Where does the term GEO come from?
Pranjal Aggarwal and five co-authors, researchers from Princeton University and IIT Delhi, posted “GEO: Generative Engine Optimization” on arXiv in November 2023. It was accepted to the KDD 2024 conference. The paper defines GEO as a black-box method for improving how visible a page is in generative engine answers, and it builds GEO-bench, a set of 10,000 queries across many domains.
“AEO” (answer engine optimization) means almost the same thing and is used in marketing blogs. Neither term is an official Google category. Google’s generative AI guide treats both as SEO and says its AI features sit on the same core ranking systems as normal Search.
What did the GEO research find?
The authors tested nine ways of rewriting content against an unchanged baseline. Here are the results the paper reports:
| Method | Result in the paper |
|---|---|
| Cite Sources, Quotation Addition, Statistics Addition | Best group: a relative gain of 30 to 40 percent on the position-adjusted word count metric (up to 40 percent in the abstract) |
| Keyword Stuffing | Weakest: scored below the baseline, and on Perplexity.ai ran 10 percent worse |
| Fifth-ranked site plus Cite Sources | Visibility up 115.1 percent, while the top-ranked site lost 30.3 percent |
| Other methods (authoritative tone, easier language, fluency, unique words, technical terms) | Weaker and mixed, varying by topic |
Read these with care. The tests ran on the authors’ own engine (gpt-3.5-turbo with the top five Google results) plus a 200-query sample on Perplexity.ai. Engines have changed since 2023, so treat the numbers as a direction, not a promise. The direction is still useful: specific, sourced, quotable content beats padded or stuffed content, and smaller sites gain more than leaders.
What does Google say about GEO?
Google’s guide (May 2026) tells site owners to skip “AEO/GEO hacks”. It says there is no special markup, llms.txt file or chunking trick for AI features, that manufactured mentions do not work, and that content created mainly to manipulate AI responses breaks the scaled content abuse policy. What it does ask for is original, expert, first-hand content on a crawlable, indexed page. Details are on the AI Overviews and AI Mode page.
How do you practise GEO, step by step?
- Fix the SEO base first. AI tools retrieve from search indexes: Google’s for AI Overviews and AI Mode, and Bing’s for much of Copilot. Indexed, crawlable, fast pages come before everything else.
- Lead with the answer. Put a 40 to 60 word direct answer under each question heading. See answer-first content.
- Add real numbers with a named source. Say “the 2024 GEO paper found…” and link it. The paper’s best methods were sources, quotes and statistics, so give models something concrete to lift. The page on statistics and quotes shows how, and never invent a figure.
- Publish what only you have. Your own survey, benchmark, pricing data or founder story cannot be found elsewhere, so engines must cite you. Start with original research.
- Show who wrote it. Author name, role, date and sources help readers and models judge trust (E-E-A-T).
- Let the right bots in. Allow search crawlers for the tools you want to appear in. See AI crawler access.
- Write for the follow-up question. AI tools split one question into several. Cover the obvious next three on the same page.
- Track it. Test your 20 to 50 questions each month and log whether you are cited (see AI visibility tracking).
How do you measure GEO results?
Three sources give real data. Google’s Generative AI performance report in Search Console shows impressions in AI Overviews and AI Mode. Bing Webmaster Tools has an AI Performance report (public preview since February 2026) with Copilot citation data. For ChatGPT and Perplexity, add a referral filter in your analytics and run manual prompt checks.
Keep a simple log: question, engine, date, cited or not, which competitor was cited. Answers vary between runs, so repeat each check and count how many times you appear instead of reading a single result.
What mistakes should you avoid?
- Stuffing keywords. The GEO paper found it the weakest tactic, and Google’s spam policies prohibit it.
- Writing for machines and forgetting readers. The citation still has to earn a click from a person.
- Chasing hacks.
llms.txt, hidden prompts and schema tricks have no support from Google (see llms.txt). - Inventing statistics to look citable. Models and readers both check, and a false figure ends trust.
- Treating one lucky answer as a ranking. Generative answers change between runs.
Where can you learn more?
Read the GEO paper itself, then work through the free digital marketing module of the Fe/male Switch startup school. Next steps: set up AI visibility tracking and write your first answer-first page.