The GEO Research Paper Explained (Princeton/IIT 2023)

"GEO: Generative Engine Optimization" by Pranjal Aggarwal and colleagues named the field and tested nine ways to make content more visible in AI answers. Adding citations, quotations and statistics improved visibility by roughly 30-40% in its tests, while keyword stuffing didn't help. Most experiments used a research engine built on GPT-3.5, so treat the results as hypotheses to test.

What is the paper, and who wrote it?

The paper was first posted to arXiv in November 2023 and published in the proceedings of KDD 2024, the ACM SIGKDD Conference on Knowledge Discovery and Data Mining, held in Barcelona in August 2024 (arXiv). The authors are:

• Pranjal Aggarwal, IIT Delhi
• Vishvak Murahari, Karthik Narasimhan and Ameet Deshpande, Princeton University
• Tanmay Rajpurohit and Ashwin Kalyan, independent researchers

The paper makes three contributions:

1. It defines generative engine optimization as a black-box framework that content creators can use to improve visibility in generative engine responses.
2. It proposes visibility metrics designed for generative engines.
3. It releases GEO-bench, a benchmark of 10,000 queries.

What problem did it address?

Search engines show a ranked list of links, so "visibility" mostly means position. Generative engines such as Perplexity, Bing Chat and Google's SGE (all named in the paper) blend several sources into one answer with inline citations. The authors argue that creators have "little to no control over when and how their content is displayed," and that generative answers may reduce the traffic websites receive. The paper asks whether creators can change their own content to be used and cited more.

How did the researchers measure visibility?

They defined two kinds of impression metrics (arXiv):

Improvement is the relative change in a source's impression after it is optimised. One source per query was chosen at random and rewritten with each method in turn.

How was the test engine set up?

Which methods did they test, and what worked?

The paper tested nine methods. Each one was applied by an LLM that rewrote the source content.

The paper reports relative improvements of 30-40% on position-adjusted word count and 15-30% on subjective impression for the top three methods. The best methods improved on the baseline by 41% and 28% on those two metrics (arXiv).

What happened on Perplexity?

On the 200-query Perplexity subset, Quotation Addition improved position-adjusted word count by 22%. Statistics Addition improved subjective impression by 37%. Keyword stuffing performed about 10% worse than the baseline on position-adjusted word count. The authors note that Perplexity doesn't let users specify source URLs, so they uploaded the source text as files instead. That is still a test setup, not ordinary use of the product.

What was the most surprising finding?

When every source in an answer was optimised, the gains went mainly to sources that ranked lower in the Google results:

The authors suggest this could help smaller creators compete. For brands that already rank first, the warning is that competitors adding evidence to their pages can take visibility away from them.

Did methods work differently by topic?

Yes. In the paper's results:

• Cite Sources worked best for factual statements and law and government.
• Quotation Addition worked best for people and society, explanations and history.
• Statistics Addition worked best for law and government, debate and opinion.
• Combining methods also helped. Fluency Optimization plus Statistics Addition beat any single method by more than 5.5% on a 200-example subset.

What are the limits?

The authors list several limitations themselves (arXiv):

Two more limits matter for marketers:

• The main engine used GPT-3.5 with five sources. Today's products use different models and retrieve many more pages.
• Content was rewritten automatically. Facts were added for the experiment, but a real page can only add statistics and quotations that are true.

How should marketers apply it?

The findings are consistent with Google's guidance on helpful, reliable content (Google):

1. Cite primary sources for the claims you make.
2. Use real numbers from your own data or from credible studies. Never invent them.
3. Quote named experts with their credentials.
4. Write fluently and simply. That helped; a pushy, persuasive tone didn't.
5. Skip keyword stuffing.
6. Test on your own prompt set with a baseline and repeated runs.

A quick example

Illustrative. A page says, "Our software saves teams a lot of time." A rewrite in the spirit of GEO says, "In our 2026 survey of 312 customers, 61% said they saved at least two hours a week (see the methods note)," and links to the survey. The numbers in this example are illustrative. The paper's own example, for the query "What is the secret of Swiss chocolate?", added a source to a consumption statistic and reports a 132.4% relative visibility improvement for that one case.

What mistakes do people make when citing the paper?

• Quoting "40%" as a guaranteed uplift for any site
• Saying it tested today's ChatGPT or Google AI Overviews (it didn't)
• Forgetting that authoritative tone and keyword stuffing didn't help
• Ignoring the finding that top-ranked sources lost visibility
• Treating one 2023 study as the whole evidence base for GEO

Checklist

• [ ] Read the paper's method and limitations sections
• [ ] Add true sources, statistics and quotations to priority pages
• [ ] Improve fluency and clarity
• [ ] Avoid keyword stuffing and inflated tone
• [ ] Test changes with a baseline and repeated runs
• [ ] Describe the findings accurately in your own content

Related

• How AI search engines build an answer
• How to write answer-first content LLMs quote

Sources

• Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, GEO: Generative Engine Optimization, KDD 2024 (arXiv:2311.09735)
• Google Search Central, Creating helpful, reliable, people-first content