The research paper that named GEO.
The companies behind ChatGPT, Gemini and Claude don’t call their products generative engines. Google calls its versions AI Overviews and AI Mode. The paper called all of them generative engines, then named a new field after that word.
Then it set out to prove that this new field needs something SEO can’t give you.
It doesn’t prove that.
What it tested was one 2023 AI model, in the researchers’ own setup, grading its own work. The quotes, statistics and sources it added could be made up - the code said so. When independent researchers tried the same methods on newer models, almost none of them helped.
This paper is not a reason to change your website.
What the name GEO is built on
The paper is “GEO: Generative Engine Optimization.” It was published in 2024 in the proceedings of KDD, a peer-reviewed data science conference.
The term came from research. A Stanford paper earlier in 2023 studied “generative search engines.” This paper dropped “search” and named the field after what was left.
A new name suggests new work. New work needs its own budget, its own tools and its own specialists. That is the case the paper set out to make.
Its premise is that “traditional SEO methods are not directly applicable to Generative Engines,” because the AI behind them “is not limited to keyword matching.”
Was SEO keyword matching in 2023? No. SEO had moved past keyword matching years before, and in 2019 Google explained how it uses language models to understand searches.
Does the paper’s own test depend on SEO? Yes. Its AI read only the top five Google results for each question. Every page in the test was on Google’s first page before anything was rewritten.
How the paper tested GEO
The researchers took 1,000 questions, most of them simple ones like “where do the hearts of palm come from.” For each question, they saved the text of the top five Google results.
Then GPT-3.5 did everything. It answered the question from those five pages. It rewrote one of the pages with quotes, statistics or sources. It answered the question again. It graded its own answers.
Visibility meant one thing: how many of the answer’s words sat in sentences that cited the rewritten page. With quotes added, that page went from 19.3% of the answer to 27.2%. That is the “up to 40%.”
No AI product people use was part of the main test. No page moved up in Google. Nobody checked whether a business got a visit, a call or a sale.
The paper says the methods also worked on Perplexity. Perplexity did not search for the pages. The researchers gave it the text “as file uploads” and had it answer only from those files.
What the researchers’ code told the AI to write
The paper says the added quotations and citations came “from credible sources.” The code the researchers published says something else:
Cite sources
“You may invent these sources but ensure they sound plausible.”
Quotation addition
Add more quotes, “even though fake and artificial.”
Statistics addition
Add “positive, compelling statistics (even if hypothetical).”
A page about Swiss chocolate picked up a survey by “The International Chocolate Consumption Research Group,” and its share of the answer went up 132.4%. A page about the Jacksonville Jaguars picked up “a testament to their prowess and determination” and no new fact, and its share went up 89.1%.
The paper’s own caption for those examples says the methods raised visibility “without adding any substantial new information.”
Can you trust what the paper found?
No, and the paper’s own numbers show why.
The score counted words. A page that got wordier, or picked up a made-up survey, got more of the answer’s words. That is what the paper measured, and that is what it found.
When the researchers rewrote all five pages for a question at once, the page ranked fifth got more of the answer and the page ranked first got less - 30.3% less with cited sources. The shares add up to 100%, so one page’s gain is another page’s loss. The paper calls this a way to “democratize the digital space.”
I would call it the same five pages, all on Google’s first page, splitting the same answer differently.
What happened when other researchers tested the same methods
In 2025, researchers at Parameter Lab, three German universities and NAVER AI Lab tested the paper’s methods again in C-SEO Bench, accepted at NeurIPS 2025. They used four newer AI models, six kinds of searches and more than 1,900 questions. They measured where the AI cited the page, not how many words it got.
Across all of those tests, the paper’s eight rewriting methods helped a page significantly once. Adding statistics moved pages later in the answer in 19 of 24 settings. On Claude 3.5 Haiku, nothing helped.
What helped in every test was being first among the pages the AI read. That is what ranking first in search gets you.
In 2026, SAGEO Arena ran the whole path - search, ranking, then the answer - and found that these rewrites “remain largely impractical under realistic conditions.”
What this means for you
Nothing in this paper tells you to change your website.
What you do is still SEO, and it covers AI answers too. Google says the same in its guide to generative AI search: optimizing for it is “still SEO.”
Adding true quotes, statistics, and sources is already part of good SEO. You do it for the person reading, not for an AI counting words.
Get your pages found first. Google’s AI Overviews and AI Mode draw on pages from Google Search, and ChatGPT runs its own search. Getting found is still the work, and a page has to rank in Google before an AI Overview links to it.
Answer what your buyers ask. A law firm, an online store and a local service business get very different questions, so they need very different pages.
Add a statistic, a quote or a source when it is true and it helps your buyer, and link to where it came from. A made-up number is a false claim with your name on it. What your website needs depends on what it has, and a study of someone else’s traffic can’t tell you that.
Measure visits, calls and leads - not the share of words a page gets in an AI answer.
Frequently asked questions
Does GEO work?
The methods in the 2023 paper raised a page’s share of one AI model’s answer in a lab test. When independent researchers tested them in 2025 on four newer models, they helped once across all those tests and often made the AI cite the page later. What gets a page into AI answers is what gets it found and makes it worth citing, which is SEO.
Is GEO different from SEO?
GEO is a newer name for the same work. Getting found in AI answers depends on getting found in search, and Google says optimizing for its AI features is still SEO. The name does not change the work.
Where did the name GEO come from?
The 2023 paper built it from “generative search engines,” a term used in a Stanford paper earlier that year. It dropped “search” and named the field generative engine optimization, on the same pattern as SEO. The companies behind ChatGPT, Claude and Gemini don’t call their products generative engines, and Google calls its versions AI Overviews and AI Mode.
Did the GEO paper test real websites?
It used the text of pages from Google’s top five results, saved once and rewritten in the researchers’ own setup. The AI that answered the questions did not visit the websites, and the rewritten versions were not published on them.
Sources
- [2311.09735] GEO: Generative Engine Optimization arxiv.org
- [2304.09848] Evaluating Verifiability in Generative Search Engines arxiv.org
- Understanding searches better than ever before blog.google
- GitHub - GEO-optim/GEO · GitHub github.com
- [2506.11097] C-SEO Bench: Does Conversational SEO Work? arxiv.org
- [2602.12187] SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization arxiv.org
- Google's Guide to Optimizing for Generative AI Features on Google Search developers.google.com
- AI Features and Your Website developers.google.com
- Introducing ChatGPT search openai.com
When you want to rank on top of Google and show up in AI answers, this is the work.
SEO + AI Search Optimization
