The 40% claim

What the paper actually is

The 40% figure comes from GEO: Generative Engine Optimization, a paper by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. It was submitted to arXiv on 16 November 2023, last revised on 28 June 2024, and accepted to KDD 2024, one of the serious data-mining conferences. This is not a vendor white paper. It is the academic work that gave the field its name.

The abstract contains the sentence everybody quotes:

GEO can boost visibility by up to 40% in generative engine responses.

Every word of that is accurate. The trouble starts when it is lifted out and set next to a screenshot of ChatGPT.

What "visibility" means in the paper, and what it does not

Alongside the method, the authors introduce GEO-bench, a benchmark of diverse user queries across multiple domains with relevant web sources attached. The uplift is measured on that benchmark. The "visibility" being boosted is a metric the authors define, based on how much of a generated response is attributable to a given source, roughly how much of the answer's substance came from you.

So the claim, stated precisely, is: under benchmark conditions, applying these content modifications increased a source's share of the generated response by up to 40% relative to unmodified content.

That is a real and interesting result. Here is what it is not:

• It is not a measurement of how often ChatGPT, Perplexity, Claude or Google AI Overviews cite you.
• It is not a 40% increase in traffic, leads, mentions or citations.
• It is not "up to 40% more visible in AI search", the phrasing that appears on agency pages.
• It is not a promise that any single tactic delivers 40%. "Up to" is doing heavy lifting, and the uplift varied by method and by domain.

The authors do not claim any of those things. The people quoting them do, usually by omission rather than assertion, which is harder to argue with and easier to forgive.

How to check this yourself in ninety seconds

1. Open arxiv.org/abs/2311.09735.
2. Read the abstract. Find the 40% sentence.
3. Now find the words "GEO-bench" in the same abstract. It is the benchmark introduced in the same breath.
4. Ask yourself where in that abstract anyone measured chatgpt.com.

That is the entire audit. It takes less time than reading most of the articles that get it wrong. We publish one detail those articles do not: the arXiv abstract page does not state how many queries GEO-bench contains. A figure of roughly ten thousand circulates widely. We could not confirm it from the abstract and have not read the full PDF, so we do not print it. That is the standard we are asking of everybody else.

Why this particular error matters

Because it is load-bearing. The 40% figure is doing a specific job in this industry: it converts a research result into a sales promise. A prospective client hears "peer-reviewed research shows a 40% uplift" and reasonably concludes that a well-understood intervention produces a well-measured outcome. Neither half of that is established.

And the field notices. Google's John Mueller has been blunt about the pattern: "The higher the urgency, and the stronger the push of new acronyms, the more likely they're just making spam and scamming." Lily Ray, who has done as much serious work here as anyone, puts it more directly: "Anybody that's pretending to be an expert in [GEO], they're lying." Those are not attacks on the paper. They are attacks on what has been built on top of it.

What we would cite instead

There is real evidence about live systems, and it is more interesting than the 40%, because it is more inconvenient:

• Ahrefs found the share of AI Overview citations that also rank in the organic top 10 fell from about 76% in July 2025 to 37.9% in January 2026, across 863,000 keyword SERPs. Graded.
• SparkToro and Gumshoe ran 2,961 prompt runs and found the same brand list repeats fewer than 1 time in 100. Graded.
• Kevin Indig analysed 3.7 million citations and found only 2.37% of cited URLs appear across all three major engines. Graded.

None of these supports a tidy percentage uplift you can put on a landing page. All three describe a system that is noisier, more fragmented and less controllable than the 40% implies. That is presumably why they get quoted less.

Questions people ask about this

Where does the 40% figure come from?

From the paper that named the field: "GEO: Generative Engine Optimization" by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, first submitted to arXiv in November 2023 and accepted to KDD 2024. The abstract states that GEO "can boost visibility by up to 40% in generative engine responses".

Is the 40% figure false?

No. The paper says it, and the paper is real, peer-reviewed and worth reading. The problem is what it is taken to mean. The uplift was measured on GEO-bench, a benchmark the authors built and introduced in the same paper, using a visibility metric based on how much of a generated response is attributable to a given source. It is a controlled result on a research benchmark, not a measurement of live citation behaviour on ChatGPT, Perplexity, Claude or Google AI Overviews.

Does that mean generative engine optimisation does not work?

It means this particular number is not evidence that it does. The paper demonstrates that content changes can shift how much a generative system draws on a source under benchmark conditions, which is genuinely interesting. It does not establish that a given tactic will get your brand cited more often by a commercial AI assistant, and the paper never claims it does.

What should I cite instead?

If you want evidence about live systems, cite studies that measured live systems, and cite their sample sizes. Ahrefs analysed 863,000 keyword SERPs and 4 million AI Overview URLs. SparkToro and Gumshoe ran 2,961 prompt runs with 600 volunteers. Kevin Indig analysed 3.7 million citations across 20,000 prompts. All three have limits, and all three are measuring the real thing.