Generative engine optimization (GEO) is a specific term coined in a 2023 research paper by researchers at Princeton University, IIT Delhi and independent collaborators, later published at KDD 2024. The paper built a 10,000-query benchmark, tested nine ways of rewriting web content, and measured which ones changed how often that content got cited inside a generative engine's synthesized answer. Three methods worked reliably - adding citations, adding quotations, adding statistics. Keyword stuffing did not. That is the actual evidence behind a term that has since been adopted, loosely, by an entire category of vendors.
"Generative engine optimization" gets used two ways online. Vendors use it as an umbrella label for anything to do with AI search visibility - functionally the same ground covered by answer engine optimization and "LLM SEO". But the phrase has a narrower, older, and more checkable origin: a specific academic paper that ran a controlled experiment and published numbers. This page is about that paper - what it tested, what it found, and how much of it still applies to how ChatGPT, Perplexity and AI Overviews work now.
Where the term actually comes from
The paper is titled "GEO: Generative Engine Optimization", first posted to arXiv in November 2023 and later accepted at KDD 2024, one of the field's major data-mining conferences. The author list includes Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, working across Princeton University, the Indian Institute of Technology Delhi, and independent researchers based in Seattle. It is the paper that gives the acronym its name - not a term a vendor trademarked, and not interchangeable in origin with "AEO", which has no equivalent single source.
What a "generative engine" is, precisely
The paper's definition matters because it is narrower than how the term gets used today. A generative engine, in the original sense, is a system that responds to a query by retrieving a handful of web sources and synthesizing an answer from them - with citations back to those sources - rather than returning a ranked list of links to click through. That description covers Google's AI Overviews and Perplexity cleanly. It fits chat assistants like ChatGPT and Claude only when they are actively retrieving and citing web content, which is not every conversation. The distinction is worth keeping: optimizing to be cited in a retrieval-and-synthesis answer is a different, more concrete target than optimizing to be "mentioned favorably" by a model drawing on training data alone.
The benchmark: GEO-BENCH
To test anything, the researchers needed a large, realistic set of queries and sources - so they built one. GEO-BENCH combines roughly 10,000 queries (split into training, validation and test sets) drawn from nine existing datasets, including MS MARCO, Natural Questions, ORCAS and LIMA, plus additional queries generated with GPT-4. The queries span 25 domains, from health and law to games and history, and are tagged along several dimensions including difficulty and intent. That range is the paper's own caveat baked into the methodology: a technique that helps in one domain does not necessarily help in another, and the paper reports domain-level variance rather than a single blanket number.
The nine methods, and what happened to each
The core experiment rewrote existing web content nine different ways and measured the change in visibility inside the generated answer, against an unmodified baseline. "Visibility" here is not a single metric - the paper tracks several, including how much of an answer's word count traces back to a given source (position-adjusted word count) and a separate subjective-impression score. The headline result across the paper is a visibility gain of up to 40% from the best-performing methods - but that ceiling was not shared evenly across all nine.
| Method | What it means in practice | Measured effect |
|---|---|---|
| Cite sources | Add citations to authoritative outside sources within the content | One of the strongest gains, especially for sources that started lower-ranked |
| Quotation addition | Add relevant quotations from experts or other sources | Large gain - roughly +41% on position-adjusted word count and +28% on subjective impression |
| Statistics addition | Add relevant statistics and numbers to back up claims | Strong, consistent gain across most domains |
| Fluency optimization | Rewrite for smoother, more natural reading flow | Moderate gain, roughly in the 15-30% range |
| Easy-to-understand language | Simplify wording and sentence structure | Positive but smaller effect than citations or quotes |
| Authoritative language | Rewrite in a more confident, expert tone | Positive mainly in specific domains such as debate and history - inconsistent elsewhere |
| Technical terms | Add domain-specific terminology and jargon | Modest, inconsistent gain |
| Unique words | Increase lexical variety / reduce repeated phrasing | Minimal effect either way |
| Keyword stuffing | Repeat target keywords more frequently | No meaningful improvement, and in several domains performed worse than the unmodified baseline |
The pattern across the table is not subtle: methods that add genuinely new, verifiable substance - a citation, a quote, a statistic - outperformed methods that only changed style or repeated keywords. Keyword stuffing, specifically, is the one method the paper's own data argues against, which is worth noting given how often "optimization" still gets read as a synonym for keyword density.
How this differs from the broader "AEO" conversation
Answer engine optimization is the umbrella term for getting cited or named across any AI answer surface - chat assistants, AI Overviews, voice assistants - and it does not trace back to one paper or one fixed methodology. GEO, in its original sense, is narrower and more testable: a specific benchmark, a specific set of rewriting methods, and published numbers attached to each one. In everyday vendor usage the two terms have collapsed into synonyms, and for a buyer comparing AI visibility tools the distinction rarely changes which product to pick. But if you are the one doing the content work rather than buying a tracker, the GEO paper is the closest thing this category has to a controlled experiment, and its results are worth treating as evidence rather than as one more marketing claim.
What still holds up, two years on
The paper's models and the current generation of assistants are not identical, and the specific percentages should be read as directional rather than as guarantees you can reproduce today. What has held up in practice, based on how citation-driven assistants like Perplexity and AI Overviews visibly work now, is the underlying logic: systems that synthesize an answer from retrieved sources reward content that hands them something concrete to lift - a number, a quote, a named source - over content that is merely confident or keyword-dense. That is consistent with what AI search optimization and LLM SEO practice both converge on independently, from a different starting point than this paper's controlled test.
Turning the findings into a checklist
Three moves come directly out of the table above, roughly in order of measured impact: cite a real outside source when you make a claim, quote an expert or a primary document instead of paraphrasing it into your own words, and back up assertions with an actual number rather than an adjective. A fourth, lower-cost move - writing in plain, fluent language rather than dense jargon - showed a smaller but real gain. What the data argues against just as clearly: do not chase keyword density as a visibility lever for generative engines. It was the one method in the study that failed to beat doing nothing.
Measuring whether any of it worked
The GEO paper measured visibility inside a controlled experiment with a fixed benchmark - a luxury most sites do not have. In practice, checking whether these changes move the needle on your own content means running a GEO audit before and after: the same set of buyer questions, asked across the same assistants, logged consistently enough to notice a change. Once you need that check repeated on a schedule rather than run by hand, that is the point where a tracked AI visibility platform starts to earn its subscription instead of a spreadsheet.