The GEO paper, explained, and the number everyone gets wrong
Nearly every GEO and AEO deck in circulation cites one paper: “GEO: Generative Engine Optimization”, published at KDD ’24. The most-repeated figure taken from it, that citing your sources delivers a +24.6% visibility lift, is not a percentage lift. It is an absolute score, from a table the authors captioned “Absolute impression metrics”, on a scale where doing nothing scores 19.5.
This page is a read of what the paper actually measured, with every number taken from the PDF rather than from someone else’s summary of it. The underlying data is published as JSON under CC BY 4.0 so you can check it or cite it.
To be clear about the target. The paper is good. It is peer-reviewed, NSF-funded, reproducible, labels its own table honestly, and states its own limitations plainly. Nothing below is a criticism of the authors. The failure is downstream, in how the number travelled.
The number everyone quotes
Table 1 of the paper reports two metrics for each optimisation method. Both are absolute scores, normalised so that the impressions of all citations in a single response sum to 1. The engine is given the top 5 sources per query, so an unoptimised source sits at roughly one fifth of the response, which is exactly why the baseline is 19.5.
“Cite Sources” scores 24.9 on Position-Adjusted Word Count and 21.9 on Subjective Impression. The figure 24.6 that circulates as a percentage is this row’s Subjective Impression sub-column. Read as a lift, it is a category error. The real relative improvement for citing sources is 27.7%.
| Method | Position-Adjusted Word Count | Subjective Impression | Real relative change |
|---|---|---|---|
| No Optimization | 19.5 | 19.3 | n/a |
| Keyword Stuffing | 17.8 | 20.2 | -8.7% |
| Unique Words | 20.7 | 20.4 | +6.2% |
| Easy-to-Understand | 22.2 | 20.5 | +13.8% |
| Authoritative | 21.8 | 22.9 | +11.8% |
| Technical Terms | 23.1 | 21.4 | +18.5% |
| Fluency Optimization | 25.1 | 21.9 | +28.7% |
| Cite Sources | 24.9 | 21.9 | +27.7% |
| Statistics Addition | 25.9 | 23.7 | +32.8% |
| Quotation Addition | 27.8 | 24.7 | +42.6% |
Four claims, checked against the paper
“Citing sources gives you a +24.6% visibility lift.”
24.6 is an ABSOLUTE score in Table 1, captioned "Absolute impression metrics", on a scale where doing nothing scores 19.5. It is not a lift. Cite Sources' real relative improvement on Position-Adjusted Word Count is 27.7%.
“GEO delivers up to a 40% visibility boost.”
That is the single best method (Quotation Addition) on the single best metric, averaged over the whole benchmark. On Perplexity.ai, the only commercially deployed engine tested, the figure is up to 37%. Most methods land between 12% and 33%, and one is negative.
“Tested across six generative engines.”
The Limitations section states: "we rigorously test our proposed methods on two generative engines, including a publicly available one." Two, not six. (Six appears in a different, later paper, arXiv 2603.29979, which is a preprint, not peer reviewed.)
“Proof that structure and citations drive AI visibility.”
Every subjective sub-metric was scored by GPT-3.5 using a G-Eval-style template, and GPT-3.5-turbo generated all the responses being scored. The headline is substantially a language model rating how prominently a language model cited a source. That is a legitimate and standard method, but it is not human-validated visibility, and no deck that quotes the number mentions it.
What the paper found that nobody quotes
GEO visibility is zero-sum by construction.
The paper normalises impressions "so that the sum of the impressions of all citations in a response equals 1", across the top 5 retrieved sources. Every point one source gains is a point the other four lose. This is not a caveat someone added later, it is how the metric is defined, and it means "everyone can raise their AI visibility" cannot be true as stated.
The gains go to whoever is losing, and the leader pays for them.
Section 5.2 tested what happens when every source optimises at once. Cite Sources produced a 115.1% visibility increase for the site ranked 5th, while "on average, the visibility of the top-ranked website decreased by 30.3%." The authors read this as democratising. For a client already ranking first, it is a warning: it is the incumbent's share that gets redistributed.
The paper's own examples add persuasion, not substance.
Table 4's caption states the methods work "Without adding any substantial new information." Its worked example for Cite Sources scores +132.4% by attributing a statistic to "The International Chocolate Consumption Research Group", a body that does not appear to exist. The authors present this as a demonstration of the method. It is also a demonstration of how the method rewards confident-sounding attribution over verified fact.
Keyword stuffing measurably hurt.
17.8 against a 19.5 baseline, an 8.7% decline, the only method to go backwards. Directly useful, and absent from every summary of this paper we could find.
Everything here was measured on GPT-3.5 in 2023.
Responses were generated by gpt-3.5-turbo at temperature 0.7, five samples, five random seeds, with the top 5 Google results as sources. The paper is careful and reproducible. But the engines it describes have been replaced wholesale since, and the authors say so: methods "may need to adapt over time as GEs evolve."
What to actually do with it
- Quotation Addition and Statistics Addition beat Cite Sources. The industry quotes the third-best method.
- Keyword stuffing is the one tactic the paper shows going backwards. Stop doing it.
- Match the method to the query type: Authoritative for debate and history, Statistics for law/government and opinion, Cite Sources for factual, Quotation for people and society.
- Fluency Optimization is the best combiner, pair it rather than running it alone.
- If you already rank first, read Section 5.2 before adopting this wholesale: the modelled gains come out of the leader's share.
- Treat the magnitudes as directional and dated. The direction is credible and replicated elsewhere; the specific percentages are GPT-3.5-era artefacts.
How this was checked
Every figure was read from arXiv:2311.09735v3 (KDD '24, Barcelona, 25-29 Aug 2024) on 2026-08-30. Table 1 was transcribed by hand; the relative percentages in the fourth column are computed from it against the 19.5 baseline and do not appear in the paper. The benchmark is 10,000 queries across 25 domains; responses were generated by gpt-3.5-turbo at temperature 0.7, 5 samples over 5 random seeds, with subjective sub-metrics scored by GPT-3.5, G-Eval-style template. The one commercially deployed engine tested was Perplexity.ai, where the best method reached 37%. Where our reading differs from the paper’s own framing, the paper is quoted directly so you can judge.
Found an error in this? Tell us and we will correct it and say what changed. That is the whole point of publishing the data.