the first real study of GEO: what actually earns AI citations

Most advice about ranking in AI answers is folklore. One piece of it is not. In 2024, researchers from Princeton, IIT Delhi, Georgia Tech and the Allen Institute for AI published the study that coined the term generative engine optimization, testing nine content tactics across a benchmark of 10,000 queries. The headline: adding quotations, statistics and cited sources lifted a page's visibility in AI answers by up to 40 percent, while old-school keyword stuffing performed worse than doing nothing at all.

The team built GEO-bench, 10,000 real queries across multiple domains, each paired with the web sources a generative engine would draw on to answer. They then modified those sources using nine tactics, from adding statistics to stuffing keywords, and measured how each change affected the source's share of the generated answer, checking results against a real deployed engine. It is the closest thing this field has to a controlled experiment.

Three tactics stood clear of the rest. Adding relevant statistics, adding quotations from credible sources, and citing sources within the content each lifted visibility on the order of 30 to 40 percent in the study's metrics. A fourth finding deserves more attention than it gets: simply improving fluency, rewriting for clearer prose with no new information added, produced gains around 28 percent. Clear writing is not a nicety. It is machine-legibility.

Keyword stuffing scored roughly 8 percent below the unmodified baseline. The tactic that defined a decade of bad SEO is actively counterproductive with language models, which parse meaning rather than count matches. If your content still carries that habit, it is now a cost.

One study, one benchmark, an engine built to mimic real ones, and results validated but not exhaustively proven across every platform. Treat the direction as strong evidence and the exact percentages as laboratory numbers rather than guarantees. What makes these findings credible is that they converge with how retrieval systems mechanically work, and nothing published since has overturned them.