What Evidence Does Princeton and IIT Delhi Research Provide for GEO Strategies?

What did the study actually test?

Each method rewrote one source's text with a language model, then measured how much more of the generated answer drew on that source.

What were the headline results?

On the test engine, the baseline position-adjusted word count was 19.5. Quotation Addition reached 27.8, Statistics Addition 25.9, and Fluency Optimization about 25. Keyword stuffing fell to 17.8, and unique words barely moved. The paper summarises its best methods as improving on the baseline by 41% on the main metric and 28% on subjective impression.

On Perplexity, the baseline was 24.1. Quotation Addition reached 29.1, a 22% gain, and Statistics Addition lifted subjective impression by 37%. Keyword stuffing performed 10% worse than doing nothing.

Do quotes from credible sources increase LLM visibility by up to 41%?

In the paper's own test engine, yes: the 41% is Quotation Addition on the main metric. But three qualifications matter:

• It is a lab result. The engine used GPT-3.5 and five sources per query, much simpler than ChatGPT, Gemini or AI Overviews today.
• The metric is share of the answer, not recommendations or traffic. A bigger share of an answer's words is not the same as being named first, or getting clicks.
• On a live engine it was smaller. On Perplexity, quotations gave 22%.

There is also a practical warning. The paper's rewrites were produced by a language model, and its examples show added statistics and quotations without substantial new information. On a real site, any quotation must be real and attributed, and any statistic must come from a source you can name. Invented quotes are a credibility and legal risk, whatever they do for visibility.

Does citing sources in your content improve generative search visibility?

In the study, citing sources, meaning naming credible sources for claims in the text, was one of the three strongest methods. Its effect depended heavily on where the source already ranked. For the site ranked fifth in search results, Cite Sources raised visibility by 115%, while the top-ranked site's visibility fell by 30% when it used it. Quotation and statistics methods showed the same pattern: large gains for lower-ranked sources, losses for the top one.

The authors present this as a way for smaller sites to compete. It also suggests the gains are partly relative: when a weaker source becomes more quotable, it takes share from the others in the same answer. For practice, the sensible reading is to support claims by naming where they come from, because it makes content more useful and easier to trust. It is not a case for adding outbound links for their own sake.

Do listicles, expert quotes, statistics, citations, and JSON-LD improve AI visibility?

The GEO paper tested three of these: quotations, statistics and citations. It did not test listicles or JSON-LD. For those, the evidence comes from elsewhere:

• Expert quotes: supported by the paper in its test setup, with the caveats above.
• Statistics: supported, and especially useful in law and government, debate and opinion queries, according to the paper's category analysis.
• Citations of sources: supported, most for facts and statements, and for law and government queries.
• Listicles: not tested in the paper. Other studies suggest comparison and "best of" pages earn more brand mentions in AI answers, as covered in whether listicles improve SEO rankings and GEO visibility.
• JSON-LD: not tested in the paper. Ahrefs tracked 1,885 pages that added JSON-LD and found no measurable gain in ChatGPT or AI Mode citations against matched controls.

The paper also found that the best method varies by subject. Quotations did best for people and society, explanation and history queries. Combining fluency improvements with statistics gave the largest gain of any pairing.

What happens when everyone uses these methods?

The paper tested one scenario where all five sources were optimised at once, and found that lower-ranked sources still gained while the top source lost ground. A later benchmark examined this more closely. C-SEO Bench, by Haritz Puerto and colleagues, published in 2025, tested conversational SEO methods across question answering and product recommendation, in six domains, with varying numbers of sites adopting them. It found most methods largely ineffective and often harmful to a document's ranking. Traditional SEO, which improves a source's position in the model's context, was significantly more effective. As more sites adopted the methods, the overall gains fell, which the authors describe as congested and zero-sum.

How should you use this research?

• Treat it as direction, not proof. It shows that content with named sources, real statistics and attributed quotes tends to be drawn on more, in controlled tests.
• Do not expect 40%. Live engines retrieve differently, use newer models and weigh many other signals.
• Get retrieved first. Both the GEO paper's rank effects and C-SEO Bench point to position in the retrieved results as a key lever. That is still mostly search visibility.
• Never fabricate. Add evidence you actually have.
• Test on your own pages. Change one thing, track citations for fixed prompts over several weeks, and compare with unchanged pages, as described in what GEO tactics can improve AI citation share and how to test them.

We have not replicated the study. The figures above come from the paper's KDD 2024 version and the C-SEO Bench paper. Other reading is listed in where to learn GEO and AEO, and Google's own view in the review of Google's GEO guide.

What might this look like in practice?

Consider an illustrative employment law firm, Harcourt Reed, whose guide to redundancy pay was rarely cited in AI answers. Instead of rewriting it with invented figures, the team added the statutory weekly pay cap from the government's own page, named the tribunal decision behind a key point, and quoted one of its partners by name. They also split a dense section into shorter paragraphs. They tracked 20 fixed prompts for eight weeks against two similar guides they left alone. The revised guide was cited in a few more answers than before, and the unchanged guides were not. The change was modest, well below the paper's headline numbers.

What do people ask about the Princeton GEO study?

Where does the "40% visibility boost" come from?

From the 2023 GEO paper by Princeton and IIT Delhi researchers, measured in a test engine built on GPT-3.5.

Did the study test ChatGPT or Google?

No. It used its own test engine and checked results on Perplexity.

Did keyword stuffing work?

No. It performed worse than doing nothing, including a 10% drop on Perplexity.

Has anyone repeated the findings?

C-SEO Bench, published in 2025, found most such methods largely ineffective, and gains shrinking as more sites adopt them.