What "GEO" Actually Means: Reading the Founding Paper So You Don't Have To

What the GEO paper actually is

Most of what circulates about generative engine optimization is second-hand: a checklist copied from a blog that copied it from another blog. The primary source is a single document. Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, and Deshpande published "GEO: Generative Engine Optimization" as an arXiv preprint (2311.09735) and then at ACM SIGKDD 2024, one of the established venues for data-mining research. The paper coined the term and, more importantly, ran a controlled test of it.

The test rested on a purpose-built benchmark the authors named GEO-bench: roughly ten thousand queries drawn from nine datasets spanning different domains and question types. For each query, the researchers took the source pages a generative engine might draw on, applied a specific content change to some of them, and then measured whether that change made the engine more likely to surface and prominently feature that source inside its generated answer.

The measurement is the part worth understanding, because it is where careful reading pays off. The authors did not measure clicks, rankings, or revenue. They designed a metric called Position-Adjusted Word Count, which approximates how much of the generated answer a given source occupies and how prominently it appears. Everything the paper claims is a claim about that visibility proxy, not about a buyer picking up the phone.

What actually moved the needle

The headline finding is refreshingly concrete. Across the benchmark, a handful of content-level interventions produced a relative lift of roughly 30 to 40 percent on the Position-Adjusted Word Count metric compared with unoptimized baselines. Not every change helped, and the ones that did were not the tricks the SEO industry might expect.

Three levers stood out consistently, and one of them was the strongest of all.

What is GEO, precisely, and what sits underneath it

It helps to be exact about what the term names. Generative engine optimization is the practice of shaping content so that a generative answer engine, the layer that reads sources and writes a synthesized response, is more likely to retrieve, feature, and cite it. The object being optimized is a citation or a mention inside an answer, not a position in a list of links.

GEO does not float in a vacuum. The engines it targets are built on retrieval-augmented generation, the architecture introduced by Lewis and colleagues in 2020, which pairs a language model with a searchable index queried at the moment a question is asked. Understanding that substrate matters, because it explains why the paper's levers work: a system that retrieves passages and synthesizes from them will naturally favor passages that are specific, sourced, and quotable. GEO is, in effect, content design for a reader that is itself a retrieval system.

GEO vs SEO: the same page, a different target

The most common misreading is that generative engine optimization is a rebrand of search engine optimization. It shares tactics with SEO, and a well-structured, credible page tends to do well in both, but the target is different. SEO optimizes for a rank in a list of links. GEO optimizes for being featured and cited inside a single synthesized answer.

That difference is not cosmetic, and it is worth stating plainly: classic ranking and AI-answer citation are measurably decoupled. A page can rank outside the top organic results and still be cited in an AI answer, and a page can rank well and go unmentioned. The founding paper does not prove this decoupling on its own, but its choice to invent a new visibility metric rather than reuse rank position is an early acknowledgment that the old yardstick does not fit the new surface. This is the intellectual basis for treating classic search and AI answers as separate things to measure, rather than one blended score.

AEO vs GEO: a terminology caveat worth keeping

You will see "answer engine optimization" (AEO) and "generative engine optimization" (GEO) used interchangeably. They are not quite the same idea, and the difference is worth holding onto even though the market has largely collapsed them together.

Industry accounts describe AEO as the older label, inherited from the mid-2010s work of optimizing content to win featured snippets ("position zero") and voice-search answers. GEO is the newer, academically coined term specific to content retrieved and cited by a generative model. They share tactics, structuring content to be extractable as a direct answer, but the underlying machinery differs: a featured snippet is an extraction from one page, while a generated answer is a synthesis across many. We flag this as an industry-consensus narrative rather than a settled, primary-sourced fact, which is exactly the kind of distinction the founding paper trains you to make.

What the paper never proved

A founding result is valuable precisely because its limits are legible. The GEO paper earns its authority by measuring one thing carefully, which means it is silent on several things people now claim it settled. Reading it carefully means holding both.

How to read a founding result

The discipline the paper models is the one worth keeping. It defined a metric before it made a claim, it reported relative lift rather than a single naked number, and it scoped its conclusions to what it measured. That is the opposite of most generative engine optimization advice in circulation, which borrows the paper's authority while dropping its caveats.

For a business, the practical takeaway is narrow and specific. The levers the study validated, credible citations, real quotations, specific statistics, are good content practice regardless of engine, because they add verifiable substance a retrieval system can use. What no study, including this one, can hand you is a guaranteed citation or a guaranteed customer. The only responsible way to know whether any of this is working for your business is to measure your actual presence in real answers to your real buyer questions, and to keep measuring as the engines change.