What Generative Engine Optimization Is, and the Paper It Came From

Generative Engine Optimization (GEO) has a definition and a research paper behind it: a 2023 study by researchers at Princeton University and IIT Delhi that gave the idea its name. You will also see it called LLM SEO or AI SEO, and you will often see it sold with one number, "up to 40%", a gain measured in a lab engine, for pages that were already among its five sources. This guide explains what the paper tested, what that number measured, and what later studies found, so you can judge the claims you hear.

AI answers in tools like ChatGPT, Perplexity AI and Google AI Overviews now sit between some shoppers and the websites they used to visit, which is why the question matters. Here is the short answer first, then the detail behind it.

The short answer

Generative Engine Optimization (GEO) is the practice of shaping your content so AI answer engines use and cite it. The term comes from a 2023 research paper, published at the KDD conference in 2024, which found that quotations, statistics and cited sources helped most in lab tests and keyword stuffing hurt.

• In the paper's lab tests, adding quotations, statistics and cited sources raised a page's share of the AI answer the most; keyword stuffing scored below doing nothing.

• The "up to 40%" figure measured how much of an AI answer drew on a page that was already one of five sources. It did not measure rankings, clicks or sales.

• Later tests on newer AI models found most of the same tactics did little. Being among the sources, and being cited by third parties, mattered more.

• Google says work for its AI search features is "still SEO", with no special AI files or markup needed.

The full definition comes next, along with the other names GEO goes by.

What generative engine optimization is

Generative Engine Optimization (GEO) is the practice of writing and structuring your content so that AI answer engines, which the paper calls "generative engines", draw on it and cite it. That covers tools such as ChatGPT, Perplexity AI, Google AI Overviews and Google Gemini.

LLM SEO, LLM optimization and AI SEO are names people use for the same goal. The LLM in those names is a large language model (LLM), the kind of AI that writes these answers. No agreed body separates the terms; GEO is the one with a founding paper.

The paper describes how an answer engine works. It searches for sources and reads a handful of them, then writes one answer and cites the sources it used. What GEO changes is your source, so that more of that answer comes from you. Search engine optimization (SEO) tries to get your page listed; GEO tries to get it used inside the answer.

The place to start is the paper that gave GEO its name.

The paper GEO came from

The term Generative Engine Optimization comes from a research paper titled "GEO: Generative Engine Optimization", first posted to arXiv, a public research archive, on 16 November 2023. It was published at KDD 2024, a peer-reviewed data-science conference held in Barcelona from 25 to 29 August 2024.

The authors argued that AI answer engines help users but take traffic and control away from the sites they draw on, and that site owners had no tested way to respond. Google Scholar listed the paper as cited by 406 other works on 22 October 2026.

The paper defines GEO on its own terms, which leaves the question of how GEO sits next to SEO and Answer Engine Optimization (AEO).

How GEO relates to SEO and AEO

In relation to SEO, GEO is a layer on top of search engine optimization, not a replacement, because AI answer engines still find their sources through search. The 2023 paper itself used the top Google Search results as its sources.

In the context of SEO, Google's own guidance treats them as one job. Google Search Central puts it this way: "optimizing for generative AI search is optimizing for the search experience, and thus still SEO" (page updated 10 July 2026). Google also says there are "no additional requirements to appear in AI Overviews or AI Mode". That guidance covers Google Search; it does not speak for ChatGPT or Perplexity AI.

On the question of what is GEO vs SEO, GEO is like SEO in that both start with useful, findable pages. It differs in that the reader may never click, because the answer is written for them. Whether GEO works at all is what the paper set out to test.

How the researchers tested it

The researchers built their own answer engine, a "generative engine" in the paper's words: it took the top five Google Search results for a question and had OpenAI's gpt-3.5-turbo write an answer citing them. They then rewrote one of those five sources using a tactic, ran the engine again, and measured how much more of the answer came from that source.

The test in brief

Each tactic is one way of rewriting the source. The nine they tested:

• Authoritative (persuasive, confident wording)

• Statistics Addition (adding figures)

• Keyword Stuffing (repeating search terms)

• Cite Sources (citing credible sources)

• Quotation Addition (adding quotes from relevant people or sources)

• Easy-to-Understand (simpler language)

• Fluency Optimization (smoother writing)

• Unique Words (unusual vocabulary)

• Technical Terms (specialist vocabulary)

The results show which of the nine tactics raised visibility most.

What raised visibility in the tests

Adding quotations did the most for a source's share of the AI answer in the paper's tests, followed by adding statistics, smoother writing and citing sources. The paper's headline, "GEO can boost visibility by up to 40% in generative engine responses", is the gain for quotations on its main measure, Position-Adjusted Word Count, for a page that was already one of the five sources, inside the researchers' own test engine. The caption of the paper's Table 1 puts the best gains at 41% on that measure and 28% on the second.

Results on the paper's main measure, calculated from its Table 1

The tactics helped sources that sat lower in the list far more than the top one. Citing sources raised the fifth-ranked source's visibility by 115.1% and lowered the top-ranked source's by 30.3% (the paper's Table 2). The authors read this as a chance for smaller sites.

What worked depended on the topic. Statistics helped most on law and government, debate and opinion questions. Quotations did best on people and society, explanation and history. Citing sources led on factual statements and on law and government, and persuasive, confident wording on debate, history and science.

On Perplexity AI, with 200 queries, the paper reports gains of "up to 37%".

One tactic made things worse, and Google lists several more that you do not need.

What did not work, and what you do not need

Keyword stuffing, the old SEO habit of repeating search terms in the text, scored 17.7 against 19.3 for making no change in the 2023 GEO paper's test engine, about 8% lower on its main measure. Google also says you need no llms.txt file, no special AI markup and no chunked AI copies of your pages to appear in its AI features.

On the paper's second measure, the AI judge's score, the method the paper calls Keyword Stuffing rose slightly (about 5%), the smallest gain of the nine. Table 1 of the paper labels it and Unique Words "Non-Performing". A 2026 review of 45 studies found that the lower score for keyword stuffing is one of the results that held up.

Google Search Central says you do not need any of these to appear in its AI features, Google AI Overviews and AI Mode:

• an llms.txt file or other "AI text files"

• Markdown copies of your pages

• breaking content "into tiny pieces"

• special schema.org markup for AI

• writing "in a specific way just for generative AI search"

Those points come from Google pages updated 10 July 2026 and 10 December 2025, and they cover Google Search only. The paper's limits matter as much as its results, starting with what it does not show.

What the paper does not show

The GEO paper shows that rewriting a page can change how much an AI answer uses it, when the page is already one of the few sources the engine reads. It did not measure rankings, clicks, traffic or sales.

• It did not test whether a page gets picked as a source in the first place.

• It did not measure clicks, traffic, rankings or sales; the authors say they did not evaluate effects on search rankings.

• Its main engine was a 2023 model, gpt-3.5-turbo, reading five sources. The authors note that engines change and that longer AI memory may change how much source order matters.

• Only two engines were tested: their own build and Perplexity AI.

• The gains are for one site improving while the others stay the same.

A review of 45 GEO studies by Olivier Martinez, posted as a single-author preprint on 15 July 2026 and not peer reviewed, put it this way: the gains are "valid within its experimental setting but conditional on a source already being present in a fixed context". It marked the 40% figure "Rejected as a general claim". The C-SEO Bench researchers add that gains shrink as more sites adopt the same tactics, a space they call "congested and zero-sum".

Newer, larger tests have since gone back to the same tactics on newer models.

What later studies found

Later studies tested the same kind of tactics on newer AI models and larger sets, and most found that rewriting a page alone did little, while being among the sources mattered more. A 2025 NeurIPS benchmark called the tactics "largely ineffective", while a University of Toronto preprint found AI answers lean heavily on third-party sources.

C-SEO Bench (2025)

Researchers from Parameter Lab and NAVER AI Lab, with university partners in Germany, presented C-SEO Bench at NeurIPS 2025. They tested eight of the 2023 paper's tactics on newer models, including gpt-4o-mini, o3 and Claude 3.5 Haiku, and found them "largely ineffective". Only 3 of 54 test cases showed a statistically significant gain, and all 3 came from two new methods the benchmark added. Moving a page to the top of the AI's source list helped far more. One limit: the models were given a fixed set of documents, not live search.

University of Toronto (2025)

A preprint from the University of Toronto, posted in September 2025, compared AI search with Google and found a "systematic and overwhelming bias towards Earned media": answers leaned on third-party reviews and publishers over brands' own pages. In US car searches, earned media made up 81.9% of AI search sources, against 45.1% for Google. One limit: it is a preprint, and some of its sorting was done by AI.

E-GEO, MIT and Columbia (2025, updated July 2026)

Researchers at MIT and Columbia University built E-GEO, a test of generative engine optimization in e-commerce: 13,747 product questions, each paired with ten Amazon listings, across five AI models. Fifteen hand-written rewriting tactics "generally did not beat a simple one-line prompt", and the "advertisement" and "storytelling" styles performed badly. What worked converged on answering the shopper's need with accurate, factual product text. One limit: it is a preprint and uses Amazon listings only.

A review of 45 studies (July 2026)

Olivier Martinez's review of 45 studies found no tactic with "a stable, longitudinal, cross-platform causal effect", and noted that "commercial claims about GEO, however, have advanced faster than the evidence". One limit: it has one author and is a preprint.

One result has held in every test: keyword stuffing does not help. The rest points the same way as Google's guidance: be a source worth citing, and be cited by others. That has plain consequences for an online store.

What this means for an online store

For an online store, including one built on Shopify, the evidence points to four steps, in this order, and none of them is a trick: keep SEO basics in place, add original figures and sources, get cited by reviews and publishers, and keep product data accurate.

1. Search engine optimization (SEO) basics stay in place, because AI answers still find their sources through search.

2. Add what others do not have: real figures, named sources and quotes that answer your buyer's question, not keyword repeats. Google asks for content people find "unique, compelling, and useful", and gives "7 Tips for First-Time Homebuyers" as an example of commodity content.

3. Be cited by others: reviews, publishers and comparison sites, since AI answers lean on third-party sources.

4. Keep your product data accurate and complete, because shopping answers read product listings, not only your blog.

The 2023 paper tested written web answers, not product listings or feeds; step 4 comes from the e-commerce study and MG Lumeo's working principle, not from the original paper.

MG Lumeo's working principle is information gain over content volume, so it works on steps 2 and 4 of generative engine optimization: original, checkable information, and the product data AI systems read. That leaves the questions readers ask next.

FAQ's

How does generative engine optimization work?

Generative Engine Optimization (GEO) works by changing the pages an AI answer engine reads, so that more of the answer it writes comes from, and cites, your page. The engine first searches and picks a few sources, then writes one answer from them; GEO starts to help only once your page is one of those sources. In the 2023 paper's tests, quotes, statistics and cited sources helped most; later tests on newer models found smaller effects, which raises the question of whether GEO replaces SEO.

Is GEO replacing SEO?

GEO is not replacing SEO; it depends on it, because AI answer engines find their sources through search, and Google says work for its AI features is "still SEO". The C-SEO Bench researchers reached the same view, calling these tactics a "complement and not a replacement" to SEO. You can check the source of all this yourself.

Can I read the GEO paper myself?

The GEO paper, "GEO: Generative Engine Optimization", is free to read on arXiv under the number 2311.09735, and the published KDD 2024 version is in the ACM Digital Library (DOI 10.1145/3637528.3671900). Start with Table 1 for the results and the section on limitations; the headline figures sit there. Below is who wrote this guide.