What the Princeton GEO study found about getting your content cited in AI answers

The study and the people behind it

The term "generative engine optimization" did not come from a marketing agency. It comes from an academic paper, GEO: Generative Engine Optimization, written by Pranjal Aggarwal (Indian Institute of Technology Delhi), Vishvak Murahari, Karthik Narasimhan and Ameet Deshpande (Princeton University), and Tanmay Rajpurohit and Ashwin Kalyan (independent researchers). It was first released on arXiv in 2023 and published in the proceedings of KDD 2024, the ACM's data-mining conference, held in Barcelona in August 2024.

The authors' starting point was simple. Generative engines such as Perplexity, Bing's Copilot or Google's AI Overviews read several sources and write one synthesised answer, usually with inline citations. A website owner no longer competes for a position in a list of ten links; they compete for how much of the answer is drawn from their page, and how prominently it is cited. The paper asks a practical question: what can a content creator change on their own page to increase that share?

How they tested it

The team built a benchmark, GEO-bench, of 10,000 queries drawn from nine sources and spanning 25 domains, from law and health to history and business. For each query they took the sources a generative engine would use, rewrote one source with a given method, and measured how its visibility in the answer changed. Visibility was scored in two ways: a position-adjusted word count (how many words in the answer are attributed to the source, weighted towards citations near the top) and a subjective impression score that rates factors such as relevance, influence and how likely a reader would be to click.

Nine rewriting methods were tested: making the text more authoritative in tone, adding statistics, keyword stuffing, citing credible sources, adding quotations, simplifying the language, improving fluency, adding unique words and adding technical terms.

What actually moved visibility

The headline result, in the authors' words:

"GEO can boost visibility by up to 40% in generative engine responses." — Aggarwal et al., GEO: Generative Engine Optimization

Underneath that headline, the detail is more useful:

• Quotation addition performed best. On the position-adjusted word count it lifted the overall score from a baseline of 19.3 to 27.2, roughly 41% better.
• Statistics addition and citing sources came next, with overall scores of 25.2 and 24.6. Fluency optimisation scored similarly (24.7).
• Keyword stuffing underperformed, scoring 17.7, below the unoptimised baseline. The tactic that defined early SEO did nothing useful here.
• A more authoritative tone on its own produced no significant improvement. Sounding confident is not the same as showing evidence.

Two further findings matter for smaller brands. First, the methods helped lower-ranked pages far more than top-ranked ones. For a source in fifth position in the underlying search results, citing sources raised visibility by 115.1%, quotation addition by 99.7% and statistics addition by 97.9%. For the source ranked first, the same methods reduced visibility by between about 21% and 30%. The paper describes GEO as especially helpful for lower-ranked websites.

Second, results varied by subject. Statistics worked best for law and government, debate and opinion queries; quotations for people and society, explanations and history; citing sources for factual statements and law; an authoritative tone for debate, history and science; and fluency for business, science and health. The authors conclude that optimisation needs to be domain-specific. They also tested the methods on Perplexity.ai, a live engine, and report visibility improvements of up to 37%.

What this means in practice

For a European SME, the study's message is encouraging: you do not need to be the biggest site in your category to be used in an AI answer, but you do need content that an AI system finds worth quoting. In our work with European SMEs we often see service and product pages that make claims without evidence. That is exactly the type of page the study suggests is easiest to improve.

1. Audit your key pages for evidence. List the ten pages that answer the questions your buyers ask most. For each, mark every claim that has no number, source or named expert behind it.
2. Add verifiable statistics. Replace "fast delivery" with your actual average delivery time; replace "widely used" with a figure and its source and year. Only use numbers you can stand behind.
3. Cite credible external sources. Link to regulators, standards bodies, research or official documentation where they support your point. This mirrors the "cite sources" method that helped lower-ranked pages most.
4. Include short, attributable quotations. A clear line from your own engineer, a customer (with permission) or a recognised authority gives an AI system a clean, attributable sentence to lift.
5. Improve readability without dumbing down. Fluency helped across business, science and health topics. Edit for clear sentences and logical flow.
6. Match the method to the topic. Legal and regulatory content benefits most from data; historical or people-focused content from quotations; factual reference pages from citations.
7. Stop keyword stuffing. Repeating the query phrase reduced visibility in the study. Write for the question, not the keyword.
8. Re-test periodically. Run your priority questions through the AI tools your customers use and note whether, and how, you are cited.

Limits and caveats

The paper is a controlled experiment, not a guarantee. The experiments start from pages that are already among the search results the engine reads; rewriting a page does not by itself get it retrieved in the first place, so crawlability and classic search visibility still matter. Generative engines also change their models and retrieval systems frequently, so precise percentages from a 2023–2024 study should be read as directional. Finally, there is an ethical boundary: "adding quotations and statistics" only works for a brand in the long run if they are accurate. Invented figures may be picked up once, but they damage trust when customers, journalists or AI systems cross-check them.

FAQ

Sources

• GEO: Generative Engine Optimization (arXiv, v3) — Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande, 2023–2024
• GEO: Generative Engine Optimization (ar5iv HTML rendering) — arXiv / ar5iv, 2023
• GEO: Generative Engine Optimization — publication record, KDD 2024 — Princeton University, 2024
• GEO-optim/GEO code and GEO-bench repository — GitHub, the authors