Citebase
Guides Pay $197

Plain definitions

What is generative engine optimization (GEO)?

Generative engine optimization (GEO) means shaping your pages so an AI system that writes answers from search results is more likely to use them and cite them. The name comes from a 2023 research paper. Today people use GEO and AEO for roughly the same work, and Google’s position is that both are still SEO.

Where the name comes from

The term was coined in a paper called GEO: Generative Engine Optimization, by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande. It was first posted to arXiv in November 2023 and revised in June 2024. The full text lists it as a KDD 2024 conference paper.

The authors define a generative engine as a search system where a language model gathers information from several sources and summarizes it into one answer. Their point was that a site owner has little control over whether, or how, their page shows up inside that answer. GEO was their name for trying to improve that.

They built a benchmark of queries across many subject areas, called GEO-bench, and tested nine ways of rewriting a source page. The abstract reports that GEO methods can boost visibility in generative engine responses by up to 40%, and that the methods that work best vary by domain.

What the paper actually tested

The nine rewrites were: sounding more authoritative, adding statistics, keyword stuffing, citing sources, adding quotations, making the text easier to understand, improving fluency, adding unique words, and adding technical terms.

The results table in the paper, on its main word-count measure:

  • Adding quotations scored best, followed by adding statistics, fluency edits, and citing sources.
  • Keyword stuffing scored below the unedited baseline. The authors note that simple SEO-style methods like it did not perform well.
  • The authors write that adding statistics and adding quotations showed strong improvements across all their metrics.

Those are useful findings. They also come with limits that most GEO sales pages leave out.

The engine was mostly their own
The main experiments used a two-step setup the authors built: take the top five Google results for a query and have gpt-3.5-turbo write the answer. They describe it as mimicking commercial engines. They also ran tests on Perplexity and report gains of up to 37% there.
Visibility is not traffic
Their main measure counts the words in the answer that cite a source, weighted by how early the citation appears. A second measure is a set of impression scores rated by a language model. Neither measures clicks, sign-ups, or sales.
The systems have changed
The experiments ran on 2023 and 2024 systems. ChatGPT, Google, Perplexity, and Copilot have all changed their search features since. None of those companies publishes the factors its live system uses to choose a source.

So the paper is evidence that some rewrites changed how often a source was quoted in a lab setup. It is not proof that the same edit will get you cited by ChatGPT next month.

GEO, AEO, and SEO side by side

Three labels. The third column is who, if anyone, publishes rules for it.
Label What people usually mean Who defines it
SEO Getting pages crawled, indexed, and ranked in search results. Each search engine documents its own rules. Google’s are in Search Central.
AEO Getting named or linked when an assistant answers a question. See what AEO means. Nobody. It is an industry label.
GEO The same goal, framed as getting your text used inside a generated answer. A research paper named it. No AI company publishes a GEO rulebook.

Google addresses both acronyms in its guide to generative AI features in Search. It says optimizing for generative AI search is optimizing for the search experience, and so still SEO. It suggests you weigh third-party AEO and GEO advice against Google’s own guidance.

The reason the three overlap is retrieval. Google says its AI features use its core ranking systems to pull pages from the Search index before the model writes. Microsoft’s Copilot documentation says Copilot sends a generated query to Bing and composes its response from what comes back. OpenAI’s ChatGPT search article says it sometimes uses other search providers. A page those searches never return cannot be quoted. That part of GEO is plain SEO.

Where GEO really is different

These are differences in what you look at, not secret ranking factors. They are observed practice, not vendor documentation.

  • The output is a sentence. A results page shows your title and snippet. A generated answer paraphrases you, and can get your price or your audience wrong. Reading what the answer says about you matters as much as whether you appear.
  • The unit is a citation, not a position. There is no rank 3 in a paragraph. The useful record is: named or not, linked or not, and which page was cited.
  • Other people’s pages carry more weight in what you see. Google’s optimization guide says its AI features can show what is said about products across the web, including blogs, videos, and forums. If a review page is the one being quoted, that page is part of your GEO whether you wrote it or not.
  • There are several engines. ChatGPT, Perplexity, Claude, Copilot, and Google each run their own retrieval. The same prompt can cite different pages on each.

One more difference is ownership. In classic SEO you mostly work on your own pages. In GEO, much of what an answer says about you may come from pages you do not control, and the answer may blend several of them into one sentence. That makes the record of what was cited as important as the record of whether you appeared.

GEO advice to keep, and advice to drop

Keep the parts that make a page better for a human reader too:

  • Add statistics you can source, with a link to where they came from. Do not invent one. A made-up number is false whether or not a model repeats it.
  • Quote people accurately and name them. A quote from your own customer with permission is fine. A quote you wrote and attributed to “a user” is not.
  • Cite primary sources, the way this page does.
  • Put a plain answer near the top. Google says you do not need to break pages into tiny chunks, and there is no ideal length. A clear first paragraph still helps the reader.

Drop these:

  • Keyword stuffing. It did worse than doing nothing in the GEO paper, and it is named in Google’s spam policies.
  • Mass-produced prompt pages. Google says creating many pages mainly to manipulate rankings breaks its scaled content abuse policy.
  • Special AI files as a ranking trick. Google says Search does not use llms.txt and that it neither helps nor harms visibility there. The llms.txt guide has the details.
  • A promised percentage. The 40% in the paper is a lab result on a benchmark. Anyone quoting it as your expected lift is misreading it.

A small GEO test you can run

  1. Pick five buyer questions and one page you think should answer them.
  2. Run the questions by hand on each engine you care about, using the manual check. Save the text and the cited links. That is your baseline.
  3. Change one thing on the page. Add a sourced statistic, or a clear sentence that says what the product is and who it is for.
  4. Wait for a recrawl. OpenAI and Perplexity each say robots.txt changes take about a day to show up, which gives a sense of their crawl timing. Google says recrawling can take days to months.
  5. Run the same questions again with the same wording. Compare what the answers say.

Expect noise. Answers vary between runs even when nothing changed. One before-and-after is a hint, not a result. If you want to know whether an edit helped, repeat the questions over several weeks and look for a pattern you would bet money on.

Related: what AEO means, and why a competitor gets named.

Sources