Beyond the Hype: What the Princeton GEO Study Actually Measured

Roth Miklós

Few acronyms have spread through marketing circles as quickly as GEO — generative engine optimization. Agencies sell it, conference talks promise it, and most commentary traces back to a single academic paper. The direct answer to what that paper established: specific, testable content changes — citations, quotations and statistics — measurably improved a source's visibility inside generative answers, by up to 40 per cent in the best case, on a simulated research engine. What it did not establish is any guaranteed uplift on a live commercial product. Reading the paper rather than the slide decks is the more useful exercise.

How the study was built

The paper, "GEO: Generative Engine Optimization" by Aggarwal and co-authors, was published on arXiv in November 2023 and presented at KDD 2024. According to the detailed walkthrough of the Princeton GEO study's method, the authors built a benchmark called GEO-bench — a large set of user queries across diverse domains — and evaluated how different content modifications affected a source's visibility inside generative-engine answers. Visibility was measured through impression-style metrics: whether a source was cited, how prominently, and how much of the answer it supported.

The companion review of what the study measured stresses the framing the authors themselves chose: a first systematic study of a new optimization problem, not a finished playbook.

Which tactics moved the numbers

The practically relevant part of the paper is the ranking of tactics. Three families of interventions performed best, as the Hungarian-language summary of the GEO findings lays out. Citing sources — adding references to authoritative external material — made content more likely to be surfaced and credited, which aligns with how retrieval-augmented systems reuse verifiable passages with attribution. Adding quotations from identifiable people or documents improved measured visibility, plausibly because quotations add extractable, attributable substance. Adding concrete, sourced statistics gave engines something specific to lift into an answer.

Just as instructive is what did not work. Classic keyword stuffing — repeating query terms — performed poorly, and in some configurations worse than doing nothing. The finding supports a conclusion many practitioners had suspected: generative engines reward informational density and verifiability, not term frequency.

Where does measurement end and interpretation begin?

An honest reading requires listing the limits. The benchmark ran against a research prototype, so absolute effect sizes may not transfer to any specific commercial engine. The paper measures visibility within answers, not downstream business outcomes such as traffic, leads or revenue. And it does not compare GEO tactics against simply having strong, authoritative content to begin with. Treating "40 per cent" as a guaranteed uplift for any website would misread the methodology.

The hype-free Hungarian guide to SME search visibility measurement adds two further qualifiers worth keeping. The authors tested the tactics largely one at a time, so combinations and long-term effects remain unexamined. And vendors sometimes cite the study as evidence of repeatable ranking gains — which is precisely what it does not claim. The guide's advice is to read the limitations section before the results section, and to use the study as a hypothesis generator: its tactics are cheap to test on your own content, and your own measurements outrank any paper.

What this means for content operations

The measured tactics have an obvious implication for how content gets produced: citations, named sources and verifiable statistics are only as strong as the verification behind them. The framework for human-in-the-loop AI content quality describes a five-stage workflow in which a human defines the brief and allowable claims, the model drafts within those constraints, and every factual claim — dates, figures, legal references — is traced to a source a person has actually opened before publication. In other words, the properties that made content more visible in the GEO experiments are the same properties a disciplined review process is designed to protect.

Measurement closes the loop. The practical guide to monitoring what chatbots say about your brand recommends a fixed prompt panel — brand questions, category questions, comparison questions — run regularly across the major assistants, with answers logged and screenshotted with dates, so a one-off oddity can be distinguished from a persistent misdescription. That is exactly the kind of evidence the GEO study's own authors call for: observation on real engines, over time.

How a marketing team should use the paper

The decision framework is straightforward. Audit your most important pages for citability: sourced claims, named experts, concrete data — or only generic statements? Treat citations as a two-way street, because pages that reference credible sources are structurally easier for generative engines to credit. Keep measurement honest: track whether your pages actually appear in AI answers for your target queries, rather than trusting a projected percentage. And demand defined conditions from anyone selling certainty. GEO, stripped of the hype, is a useful discipline with a promising early evidence base — and defined conditions are what evidence looks like.