The Princeton GEO paper, "GEO: Generative Engine Optimization" by Aggarwal and colleagues, tested nine changes to webpage content and measured how each affected the page's visibility when a generative engine answered a question. Its headline finding is that adding citations, quotations from credible sources, and statistics could lift a page's visibility in generative answers by more than 40 percent across various queries, while keyword stuffing, a traditional SEO tactic, produced a negative effect.
In short
- The paper named the field of GEO and gave it an early evidence base, presented at the KDD conference in 2024.
- It tested nine content edits against a benchmark of queries and measured visibility in generated answers.
- Evidence-based edits won: adding citations, quotations, and statistics produced the largest lifts.
- Keyword stuffing hurt, and several widely quoted per-tactic percentages should be treated with care.
What did the study set out to test?
The study asked a practical question: if generative engines now answer queries by composing responses from sources, what changes to a source page make it more likely to be surfaced and featured in those answers? To find out, the researchers defined a set of nine content optimization methods and tested each one.
The nine methods spanned both substance and style. They included adding citations to sources, adding quotations from credible sources, adding statistics, improving fluency and readability, adopting a more authoritative tone, and, as a deliberate contrast, keyword stuffing carried over from traditional SEO. The point was to compare evidence-based and credibility-based edits against the keyword-focused habits of the older search era.
How did it measure visibility in AI answers?
The researchers built a benchmark of queries spanning many topics and used it to test the methods systematically, rather than relying on a handful of examples. They then measured each edited page's visibility in the generated answer using a metric that accounts for both whether the source appeared and how prominently, sometimes described as a position-adjusted measure of how much of the source made it into the response.
The main experiments used a generative engine setup built on a large language model, and the researchers also validated key results against a live generative search product. That combination, a controlled benchmark plus a check against a real product, is part of why the paper is taken seriously as an early, credible signal rather than an anecdote.
What did it find?
The clearest result is that evidence and credibility beat keywords. The methods that added citations, quotations from credible sources, and statistics produced the largest improvements in visibility, with the paper reporting lifts of more than 40 percent across various queries for the best-performing changes, and improvements of up to around 37 percent when validated on a live generative engine. The paper summarized that several of its nine tactics boosted visibility by roughly 30 to 40 percent.
Stylistic improvements helped too, though by less: making content more fluent and readable, and adopting a more authoritative tone, both improved visibility, but not as much as adding hard evidence. The standout negative result is keyword stuffing, which reduced visibility. That single contrast is the paper's most quoted practical lesson: the tactics that made content more credible and better-evidenced worked, while the tactic carried over unchanged from old-school SEO backfired.
What should you be careful about?
The paper is a strong early signal, but it should be read with a few caveats so you do not overstate it.
First, several precise per-tactic percentages that circulate online, attributing an exact figure to each individual method, are cleaner than what the paper itself cleanly states. The defensible takeaways are the directional one, evidence-based edits produced the largest lifts, and the headline figure, improvements of more than 40 percent for the best changes, rather than a tidy ranked list of exact per-method numbers. Second, it is one study on a specific testbed, and generative engines have changed considerably since it was run, so the magnitudes are a guide, not a guarantee. Third, like most work in this area, it measures association under controlled conditions; real-world results depend on your content, your competitors, and the platform.
None of this undermines the core lesson. It just means treating the paper as a well-supported direction, adding evidence and structure helps, keyword stuffing hurts, rather than as a precise formula.
What does it mean for your GEO work?
The practical implication is direct: build credibility and evidence into your content, and drop the keyword habits that no longer help. Concretely, that means front-loading direct answers, supporting claims with specific statistics, quoting credible sources, citing your sources, and improving readability, the changes the paper associated with higher visibility. It also means not stuffing keywords, which the paper found actively reduced visibility.
These content moves are the on-page half of GEO, and they pair with the off-page brand presence work that other research highlights as the strongest correlate of AI visibility. For how to apply the specific evidence tactics, see how statistics, quotations, and citations boost AI visibility, and for the full picture see the GEO complete guide.