What is generative engine optimization (GEO)? What the research shows and what Google says
Generative engine optimization (GEO) means preparing your content so that generative search systems use it as a source, mention it and link to it. That covers ChatGPT search, Perplexity, Gemini, and Google’s AI Overviews and AI Mode. The term comes from a 2023 study. In short, GEO is not a replacement for SEO. It is SEO with a new thing to measure: instead of only asking where you rank, you ask whether an AI answer mentions you.
Many articles about GEO come from vendors selling GEO packages. Here I stick to two first-hand sources: the original study and Google’s own documentation.
Where the term comes from
The term goes back to the paper “GEO: Generative Engine Optimization” by Pranjal Aggarwal and colleagues, published in November 2023 and accepted at the KDD 2024 conference. For it, the authors built GEO-bench, a dataset of 10,000 search queries from different subject areas, and tested which changes to a source’s text increase its visibility in generated answers. Their result: up to 40% more visibility.
What worked in the study
The researchers compared nine methods, each applied to the same original text. One of the things they measured was how much of a source’s text ends up in the answer, weighted by position.
| Method | Result in the study |
|---|---|
| Adding quotations, adding statistics, citing sources | 30 to 40% more visibility, the best methods |
| More fluent, easier-to-understand wording | 15 to 30% more visibility |
| Authoritative, persuasive tone | mixed: no significant effect according to the text, slightly above baseline in the table |
| Keyword stuffing | worse than the unchanged text |
Two details matter more in practice than the percentages:
- Smaller sites gain the most. A source ranked fifth in regular search gained 115% visibility with the “cite sources” method. The source ranked first lost 30% on average with the same method.
- The effect depends on the topic. Which method works best differs by subject area. The authors therefore recommend testing separately for each domain.
One caveat: most tests ran on a rebuilt system based on GPT-3.5. On Perplexity, with 200 queries, adding quotations brought 22%. Today’s systems work differently. Take the direction from the study, not the exact numbers: verifiable facts, sources and clear language help, keyword stuffing hurts.
What Google says about GEO
Google doesn’t use the term GEO, but its documentation “AI features and your website” describes how pages get into AI Overviews and AI Mode. The page was last updated on 10 December 2025. The key points:
- There are no additional requirements and no special optimizations needed to appear there. SEO best practices still apply.
- A page has to be indexed and eligible to be shown with a snippet to be linked as a source. A
nosnippetrule therefore excludes it. - You don’t need new machine-readable files, AI text files or special markup.
- Both features may use a technique called query fan-out: they run several related searches on subtopics internally and build the answer from those results.
AI Overviews have been available in Germany since 26 March 2025, in German and English. For Google, then, GEO is mostly good SEO. ChatGPT, Perplexity and the rest have no comparable documentation, but they also draw on the open web and need pages they can fetch.
GEO, LLMO, AEO: the terms
- GEO (generative engine optimization): the term from the study and the one most German-language articles use. “GEO SEO” means the same thing.
- LLMO (large language model optimization): visibility in language models, in practice a synonym.
- AEO (answer engine optimization): an older term from the days of featured snippets and voice assistants, now often used to mean the same thing.
All three come down to the same question: does the answer mention you or not?
What you can do
- Let AI crawlers in. Some CDNs and firewalls now block AI bots by default. The free AI Crawler Check shows which bots your robots.txt allows. How to control individual bots is covered in AI crawler controls in robots.txt.
- Make sure pages are indexed and can show a snippet. No index and no snippet means no source link in Google’s AI features. Check important pages for
noindex,nosnippetand canonical errors, for example with the Canonical Checker. - Make facts verifiable. Numbers with a source, quotes from real people, data with a date. That is exactly what helped most in the study.
- Answer the question directly and cover the subtopics. The first paragraph should contain the answer. Because of query fan-out, the obvious follow-up questions count too.
- Get mentioned on third-party sources. AI answers rely on sources they trust: trade media, directories, comparison sites, forums.
- Measure before you optimize. Search Console, Bing Webmaster Tools and GA4 now provide first-party AI data. What they show and what they don’t is covered in Measuring AI search visibility.
What you can skip
- llms.txt as a ranking factor. The file does no harm and is quick to build, for example with the llms.txt Generator. But Google says explicitly that no AI text files are needed, and no major engine has committed to reading it.
- Keyword stuffing. In the study, the only method that did worse than changing nothing.
- Mass-produced AI text to “feed the models”. Google’s spam policies apply here too, and thin pages are not a source anyone wants to cite.
Do you need a GEO agency?
Not necessarily. Most of the levers are technical SEO basics and good content, which your team can implement itself with a clear priority list. An agency makes sense if you need ongoing content, PR and link building at scale. If you first want to know where you stand in AI answers and why, an audit is enough. How I do that is on the AI search consulting page.