The Generative Engine Optimization Paper: What Aggarwal et al. (2024) Actually Proved

What the generative engine optimization paper set out to do

The paper, formally "GEO: Generative Engine Optimization" by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande, appeared in the Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining in 2024. KDD is one of the most selective venues in data mining, which is the first thing that separates this work from a marketing blog: it survived peer review at a top conference.

Its starting observation is structural. A classic search engine returns an ordered list of links and leaves the selection to the reader. A generative engine reads across sources and returns a single synthesized answer with a handful of sources named inside it. The authors argue this is a genuinely new response category, not a cosmetic feature, because the unit a business competes for is no longer a rank in a list but a place inside the composed answer.

From that observation the paper poses a researchable question: if the object being targeted has changed, can we measure a source's presence inside a generated answer, and can we identify content changes that reliably increase it? That framing is what makes the work testable rather than anecdotal.

What is generative engine optimization, in the paper's own terms

In common usage, generative engine optimization now names a broad set of tactics for being cited by AI answers. In the paper it means something narrower and more precise: a method for adjusting content so that it is more visible inside the answer a generative engine composes, measured against a defined benchmark rather than asserted.

The distinction is worth holding onto. A vendor uses the phrase to describe a service. The authors use it to describe a measurable optimization problem with an input, the content, and an output, the source's visibility in the synthesized response. When this article says the GEO paper proved something, it means proved in that second sense: shown to move a defined metric on a shared benchmark, under stated conditions.

The benchmark: how the researchers measured visibility inside an answer

The paper's most durable contribution is an instrument, not a single tactic. The authors built a benchmark to evaluate generative engine responses across a range of real user queries, and defined visibility metrics that capture how prominently and how substantively a given source appears inside the composed answer, rather than simply whether a link was returned.

This is the part that raises GEO above the level of opinion. A benchmark lets a claim be checked. If an intervention is said to help, the benchmark makes it possible to run the query with and without the change and observe the difference in the source's measured visibility. That is the ordinary machinery of empirical science, applied for the first time to the question of being named inside an AI answer.

It is also why the paper is a better foundation than the volume of "GEO tips" that followed it. Most of that trade material offers tactics with no measurement attached. The academic work offers a way to measure, which is the thing that lets any later claim be tested rather than believed.

The tested interventions: what actually moved a source's visibility

Against that benchmark the authors tested a set of content interventions to see which ones changed a source's visibility inside generated answers. The interventions that measurably helped in the systems they tested were credibility signals in the content itself: adding cited statistics, including relevant quotations, and referencing authoritative sources.

The direction of that result is intuitive once stated, but the value is that it was measured rather than assumed. It points at a specific mechanism. Generative engines appear to favor sources whose content carries the markers of substantiation, numbers with provenance, quotable expert language, and citations, when composing an answer. That is a different profile from the link-and-keyword signals that classic ranking rewarded.

GEO vs SEO: what changes when the object is a citation, not a rank

The clearest way to see what the paper implies is to set it against the model it succeeds. In 1998, Sergey Brin and Lawrence Page published the PageRank paper, which reframed relevance as a link-based vote of confidence and became the foundation of algorithmic search. In that world the object of pursuit is a position in an ordered list, and the currency is links.

The GEO paper describes a different object. When an engine synthesizes an answer, there is no list to hold a position in. A source is either drawn into the composed answer or it is not, and the currency shifts toward the substantiation signals the study measured. This is the substance behind the common shorthand of GEO vs SEO: a change in what is being pursued, from being ranked among links to being cited inside an answer, not a rebrand of the same work.

Neither model replaces the other outright. Classic ranking still governs the link results that persist alongside AI answers, and the two share underlying signals such as authority and clarity. The paper's point is narrower and more useful: the generative case has its own measurable dynamics, and treating it as identical to ranking leaves those dynamics unaddressed.

Answer engine optimization and GEO: same shift, different labels

The market uses several names for the discipline this paper helped launch. Answer engine optimization, often written AEO, and generative engine optimization are frequently used interchangeably, and the difference between AEO and GEO is more a matter of emphasis than of substance. Both describe working to be the source an engine names when it answers a question rather than the link a user clicks.

The terminology is still settling, and that is itself worth stating plainly rather than papering over with a confident definition. What the academic literature contributes is a fixed method, not a fixed vocabulary: a benchmark and a set of measured interventions that any of these labels can be held to. When the label matters less than the measurement, the field has matured a step.

The real limits: what the GEO paper does not prove

An evidence-first reading has to be as clear about the ceiling as the floor. The GEO paper is a single study, however rigorous, and generative engines change underneath it as models are updated. Effects measured in 2024 are not guaranteed to hold on a 2026 system, and the paper never claimed otherwise.

The surrounding market data reinforces the caution. Claims about how large the AI-search shift already is remain genuinely contested: reported figures for how much of search behavior AI now commands vary by roughly five to ten times depending on the vendor and, crucially, on which denominator is being counted. Any single "AI search share" statistic should be read as provisional until the denominator is named.

The forecasting record is a further reason for humility. In February 2024, Gartner predicted that traditional search-engine volume would drop 25 percent by 2026 as AI chatbots absorbed queries. As of this writing that collapse has not materialized as stated, and Google still holds the large majority of the search market. The GEO paper is strong precisely because it measures a specific effect rather than forecasting a revolution, and it should be used for what it demonstrates, not stretched to underwrite predictions it never made.

Why this grounds an AI-answers measurement in evidence, not assertion

The practical value of the paper is that it moves the AI-answers question from belief to measurement. Being present inside a synthesized answer is not a vanity concern. Pew Research Center tracked the browsing of 900 consenting US adults across 68,879 Google searches in March 2025 and found that users clicked through to a traditional result in about 8 percent of searches that showed an AI summary, against 15 percent without one. When the answer is composed, the click that used to reward a good ranking is often never made.

That is the same pressure the long rise of zero-click search has applied for years, with independent clickstream analysis putting the no-click share of US Google searches above two thirds by early 2026. In an environment where the answer increasingly resolves on the surface itself, whether a business is named inside that answer is a real commercial fact, and it deserves to be measured with the same rigor the GEO paper brought to visibility in the lab.

This is where the academic literature does its quiet work. A measurement of AI-answer visibility that rests on a peer-reviewed benchmark and a set of tested interventions is standing on evidence, not on a vendor's claim about what the engines want. That is the difference between a number you can defend and a number you have to trust.