In 2024, researchers at Princeton, Georgia Tech, and IIT Delhi rewrote web pages nine different ways and tested which versions an AI answer engine actually lifted into its responses. Adding direct quotations produced the biggest single gain, a 41 percent lift over an unoptimized baseline. Statistics addition came in close behind at 31 percent. The paper, GEO: Generative Engine Optimization, is now the most cited research behind the entire GEO industry, and also one of the most casually misquoted. Writers repeat its headline numbers constantly and rarely check what was actually tested, on which model, or how much of it has held up since.
What the original GEO study actually measured
Three researchers built GEO-bench, a dataset of 10,000 real and GPT-4-generated search queries, each paired with the cleaned text of Google's top five results. They tested nine rewrites against gpt-3.5-turbo acting as the answer engine, averaged results over five random seeds, and confirmed the pattern held on Perplexity's live search product too. This piece assumes you already know what GEO means and skips straight to the data; if you need the baseline definition first, we've covered that separately.
The metric behind every number in the paper is Position-Adjusted Word Count. It counts how many words from a given source survive into the AI-generated answer, weighted by where those words land. A source supplying the opening sentence of an answer scores far higher than one contributing a clause near the end, which is the researchers' way of saying that getting cited and getting cited prominently are not the same achievement.
Position-Adjusted Word Count rewards a source for showing up early in an AI's answer, not merely for showing up at all.
The tactics that worked, ranked by how much they moved the score
Against a baseline score of 19.3, four rewrites produced double-digit gains and five did not. Quotation addition and statistics addition led the field by a wide margin, followed closely by fluency optimization and citing outside sources.
Quotation addition: 27.2, about 41% above baseline
Statistics addition: 25.2, about 31% above baseline
Fluency optimization: 24.7, about 28% above baseline
Citing outside sources: 24.6, about 27% above baseline
Using technical terms: 22.7, about 18% above baseline
Writing in easy-to-understand language: 22.0, about 14% above baseline
Adopting an authoritative tone: 21.3, about 10% above baseline
Swapping in unique words: 20.5, about 6% above baseline
Keyword stuffing: 17.7, about 8% below baseline
Quotation addition and statistics addition likely won for the same practical reason. Both give a language model something concrete and attributable to lift almost verbatim, a number or a named speaker, without the model having to invent or paraphrase a fact it could get wrong. Fluency optimization placed third for a related reason: simplifying a sentence makes it easier to extract cleanly, independent of how much verifiable information it carries.
Why citing sources alone barely moved the needle
Citing outside sources ranked fourth in the test, a 27 percent gain, meaningfully behind quotations and statistics despite being the tactic most GEO advice treats as foundational. The paper's own headline figure, up to 40 percent, lines up with quotation addition's result rather than with citations on their own.
That ordering is a useful corrective for anyone building a GEO checklist. Citing a source is good practice, and probably helps with trust signals the paper didn't measure directly, but it isn't the single highest-impact rewrite available on its own. A page that cites a source in otherwise generic marketing language gained less in this test than a page that simply quoted someone by name or stated a specific number.
What's held up since 2024, and what's still unproven
A November 2025 analysis of 129,000 domains and over 216,000 pages, run by SE Ranking and reported by Search Engine Journal, found pages with 19 or more statistical data points averaged 5.4 ChatGPT citations, against 2.8 for data-light pages. That's a real-world echo of the lab result, a year later and on a live platform rather than a controlled test.
That pairs with the topical-authority research VizibleAI has covered before, which found how much of a topic a site actually covers predicts AI citations far better than domain rank does, a separate but related argument against generic, thinly sourced content. Researchers at Carnegie Mellon followed up again in October 2025 with AutoGEO, a system that has AI models explain their own citation preferences and turns those explanations into training signals for automated rewriting. The project treats GEO-bench as a continuing baseline rather than a dated artifact, itself a sign the 2024 findings outlasted the specific models they were tested on.
None of this means statistics and quotations behave identically on every engine today. The original study measured gpt-3.5-turbo and Perplexity specifically, both long since superseded by newer models and retrieval pipelines. The exact percentages are best read as proof that the pattern exists, not as a conversion rate that transfers unchanged to ChatGPT, Gemini, or Claude in 2026.
What doesn't work: keyword stuffing and padding word count
Two tactics failed outright in the same test. Keyword stuffing scored below the baseline, meaning it actively hurt a page's chances rather than helping them. Simply adding more words without adding new information produced no measurable gain at all.
A December 2025 Ahrefs study of 174,048 pages found the same thing from a different angle, as VizibleAI covered when that data first came out: a Spearman correlation of just 0.04 between word count and getting cited in Google's AI Overviews, a figure the study itself called a near-zero relationship. The same dataset, drawn from 560,346 AI Overviews, found the correlation between word count and citation position was identically close to zero, undercutting the idea that trimming or padding a page changes its odds either way.
Frequently Asked Questions
What is the GEO-bench study, and who ran it?
GEO-bench is the benchmark Princeton, Georgia Tech, and IIT Delhi researchers built for the 2024 paper GEO: Generative Engine Optimization, presented at KDD 2024. It pairs 10,000 real and synthetic search queries with the cleaned text of Google's top five results for each one, then tests content rewrites against an AI answer engine to see which ones get quoted more often.
Does adding statistics to a page really increase AI citations?
In the original study, adding statistics lifted the Position-Adjusted Word Count score by about 31 percent over an unoptimized baseline, the second-largest gain of nine tactics tested. A separate November 2025 analysis of 129,000 domains found pages with 19 or more data points averaged nearly double the ChatGPT citations of data-light pages, though that figure is correlational rather than a controlled experiment.
Which GEO tactic worked best in the research?
Adding direct quotations produced the largest single gain, about 41 percent over baseline, ahead of statistics addition, fluency optimization, and citing outside sources. Quotations likely help because a model can lift an attributed quote into its own answer with less risk of misstating a fact than it faces when paraphrasing a number.
Does keyword stuffing help with AI search visibility?
No. In the same 2024 test, keyword stuffing scored below the unoptimized baseline, meaning it made the page perform worse, not better. Simply adding more words without adding new information also produced no measurable gain, a pattern a December 2025 Ahrefs study of over 174,000 pages later confirmed for word count generally.
Do these findings still apply to ChatGPT, Gemini, and Claude in 2026?
The original study tested gpt-3.5-turbo and Perplexity's live search product, both since superseded. An October 2025 follow-up project, AutoGEO, found AI models' citation preferences could still be extracted and taught to a rewriting system, suggesting the underlying pattern persists even as individual models change, though the exact 2024 percentages shouldn't be treated as fixed targets for every engine today.
What is Position-Adjusted Word Count?
Position-Adjusted Word Count is the metric the GEO-bench researchers used to score each tactic. It counts how many words from a source appear in an AI-generated answer, weighted by how early those words appear, so a source quoted in the opening sentence scores higher than one buried near the end.