In short
- Five, with their limit
- Every study states what it does not say. In research that is not modesty but precision.
- With a real link
- Every source has its address. A reference without a destination is a name and not a source.
- Not ours
- This is other people's research. Our own measurements sit apart, with the date they were run.
Which eleven studies are here, and what do they show?
They are listed below with their core and their address, in an order that is not chronological but increasing in usefulness: from what is established about content, through what is established about measuring, to what is established about the relation with classic findability.
| Study | What it says | The source |
|---|---|---|
| Who checks the AI that writes your copy? | A second model, from a different family, that is not allowed to know who wrote the text. | Pombal, Rei & Martins (arXiv 2604.06996, 8 April 2026) · Yang et al. (arXiv 2604.22891, 24 April 2026) · Saraf et al. (arXiv 2508.21164, 28 August 2025). All three figures were found back on 11 August 2026 in the abstract of the paper itself. |
| How do you know whether a fix worked? | Only by measuring the same questions on the same engines again, and knowing what the platform itself changed in the meantime. | seoClarity (Mitul Gandhi, May 2026, updated June 2026) · SISTRIX (Johannes Beus, 1 May 2026). Both figures were found back in the source text itself on 10 August 2026. |
| Can you measure a position in an AI answer? | No. | SparkToro (Fishkin & O'Donnell, 2026) · PetraLabs (Tom Zu, June 2026) · MaxAEO (Chris Han, July 2026). Every figure on this page was found back in the source text itself on 10 August 2026. |
| How do you measure brand visibility across several AI engines at once? | The study describes how to measure brand visibility in AI answers systematically and at scale across several engines, with a methodology that is repeatable — the same questions, the same treatment, so that outcomes stay comparable between brands and sectors. | (GEO research) (2026). Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines. arXiv 2606.20065 |
| How often should you measure AI visibility? | The research repeated measurements of brand visibility in AI answers and analysed the variation between runs. | (GEO research) (2026). Don't Measure Once: Measuring Visibility in AI Search. arXiv 2604.07585 |
| Does GEO replace SEO, or does it come on top? | Kargaev sets SEO ranking factors and GEO factors side by side in a normalised framework with a divergence index. | Kargaev (2026). The SEO-to-GEO Gap. SSRN, abstract 6476021 |
| Where does AI get its picture of your company from? | Chen et al. | Chen et al. (2025). Generative Engine Optimization: How to Dominate AI Search. arXiv 2509.08919 |
| Which changes raise your visibility in AI answers? | Aggarwal et al. | Aggarwal et al. (2024). GEO: Generative Engine Optimization. KDD '24. arxiv.org/abs/2311.09735 |
| How much of the AI market does a measurement cover? | That depends entirely on how many engines you measure, and the gap is wider than most figures suggest. | Our own calculation. The engines come from the measurement's core list, the shares from Statcounter · AI chatbot market share (July 2026). Every percentage on this page is computed at build time and appears nowhere as text |
| Why did it say 31 at the top and 14 below? | Because they were two different measurements that looked like one. | Our own case. The three figures and the reason sit in the comment above the function that calculates the headline score |
| Did the Gartner prediction about search traffic come true? | Not as it was worded. | Gartner, press release of 19 February 2024, from “Predicts 2024: How GenAI Will Reshape Tech Marketing” · analyst Alan Antin. Checked at the source on 10 August 2026. What actually happened appears here without a counter-figure, with the reason attached. |
How do I recognise a GEO figure that is being misquoted?
By three things that are nearly always missing together: the scope of what was studied, the period, and the question the researchers actually asked. A percentage doing the rounds without those three usually comes from a study that answered a different question — and then the figure is correct and the conclusion is not.
The costliest variant is a figure about click behaviour used as a figure about mentions. Those are two different things: whether someone clicks through after an answer, and whether you were in that answer. The first depends on the user, the second on the engine.
Frequently asked questions
Were these studies paid for by you?
None of them. They are public third-party publications, with their address alongside so you can read them without us. That is why they are here rather than a summary we wrote ourselves.
What if one of these studies becomes outdated?
Then it stays with the caveat attached rather than being removed. A study contradicted by a newer one is itself a fact about how fast this field shifts, and that speed is more useful to a reader than a tidy list.
Read on
Free, no account and no card.