Skip to content
Contact

WritingGuide

Where does GEO's “up to 40%” come from? What the paper says

By · Published · Updated

Sources checked on

Many articles about GEO repeat one sentence: "boosts visibility by up to 40%". The number does come from a paper. This guide covers what the paper measured, what "up to" means, and what to check before you carry the figure over to your own site. We opened the paper's abstract page and read the full text on 25 October 2026.

Which paper does the 40% come from?

The paper is titled "GEO: Generative Engine Optimization". Its authors are Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. It was submitted to arXiv on 16 November 2023. Version 3 is dated 28 June 2024, and the arXiv page carries a "KDD 2024" comment.

The abstract says: "Through rigorous evaluation, we demonstrate that GEO can boost visibility by up to 40% in generative engine responses." The figure comes with "up to" attached. The introduction also speaks of an increase of "over 40%" for the best methods. According to the caption of Table 1, the best methods improve on the baseline by 41% on the first metric and 28% on the second.

What did the paper measure?

The authors did not test a live product. They built their own setup:

  • Queries: GEO-bench, a set of 10,000 queries. Each query comes with the cleaned text of the top 5 Google results.
  • Answer generation: the gpt-3.5-turbo model writes the answer from those 5 sources.
  • Intervention: for each query, one of the 5 sources is picked at random and rewritten with a GEO method. The other four stay as they were.
  • Metric: "visibility" is defined two ways. The first counts the words in sentences that rely on the source and where they sit in the answer (Position-Adjusted Word Count). The second is a subjective impression score that GPT-3.5 gives on seven sub-criteria (Subjective Impression). Both are normalized so that the shares of all cited sources in one answer add up to 1.

Visibility in the paper is a source's share inside one answer. Visits, clicks and sales do not enter the measure.

Table 1 compares nine methods. The best performers were adding quotations (Quotation Addition), adding statistics (Statistics Addition) and citing sources (Cite Sources). Keyword stuffing, familiar from classic SEO, did not do well. The authors also report that the effect of a method varies by domain. Citing sources, for example, stands out on factual questions.

They also ran 200 examples through Perplexity.ai. Because Perplexity does not accept source URLs, they uploaded the source text as files. Quotation Addition came out 22% above the baseline on the first metric. Cite Sources and Statistics Addition improved the two metrics by up to 9% and 37%. Keyword stuffing did 10% worse than the baseline.

What does "up to" mean?

"Up to 40%" does not mean every site gains 40%. It is the ceiling reached by the best method, on the best metric, in this setup. The same paper says the effect of methods differs by domain. When a sales page repeats the claim, "up to" tends to drop out and only "40% more visibility" remains.

Carrying the number to your site

The points below are our reading. The paper does not state them.

  1. The setup. A single model, gpt-3.5-turbo, writes the answer from 5 sources. The authors say this resembles how commercial engines work. We have not verified that independently. The paper contains nothing showing that today's products would give the same result.
  2. The metric. A bigger share does not mean more visits. The paper does not show that more people click when your share rises.
  3. One-sided change. The four competing sources stay untouched. In real life everyone edits their pages.
  4. Google's view. Google's guide to optimizing for generative AI search (last updated 10 July 2026) notes that people call this work "GEO" and "AEO", and says that from Google Search's perspective it is still SEO. It also tells site owners to review any third-party GEO advice or service.

So what should you do?

The methods the paper rates highest are good writing habits anyway: cite your sources, quote accurately, add figures that can be checked. Do those things, but never invent a number. An unsourced statistic hurts readers and your chances of being cited. If a service quote promises "40% more AI visibility", ask which metric and which site it was measured on. For how GEO differs from SEO, see What is GEO?. For Google's own position, read our guide to Google's generative AI advice.

Frequently asked questions

Does GEO really increase AI visibility by 40%?

In the paper's own test setup, some methods raised a source's share of an answer by up to about 40%. The paper does not show that a real site would get the same result.

40% of what?

A relative increase in one source's share of one answer, compared with its share before the change. It is not a rate computed on visitors or sales.

Which methods came out best?

Adding quotations, adding statistics and citing sources. Keyword stuffing fell below the baseline. The ranking comes from the test setup, so test it on your own site.

Where can I read the paper?

On arXiv as 2311.09735. The link is in the sources below.

Sources

All read on 25 October 2026.

Want to talk about this for your brand?

Write about your project and what you want to do. The message goes straight to Rues.

Open the contact form