How to read an AI visibility study

You are going to be shown studies. Every vendor in this market publishes them, we publish them, and most of them are produced by companies with a commercial interest in the conclusion. That is not enough to throw them out: some of the best data in this field comes from the people selling, because they are the only ones who have it. It does mean you need a way to read them.

In July 2026 a methodological survey reviewed 45 studies of generative engine optimization published between November 2023 and July 2026. Its criticisms of the literature make an unusually good checklist, and we have used it on our own work.

1. Does it assume you were already found?

The most common flaw. A study puts five pages in front of a model, changes one of them, and measures which gets cited. That answers “once you were found, were you chosen?” and skips the harder question: whether the model would have found you in the first place.

It matters, because the two can pull in opposite directions. One test across 171,003 pages found that working only on the text of a page pushed it out of the top twenty results about one time in ten, and the citations fell with it. The very moves that got a page picked, once it was in the running, were knocking it out of the running. Winning the selection is worth nothing if you never qualified.

Ask: did the model find that page by itself, or was it handed to it?

2. Who is judging?

A recurring pattern: the same family of models writes the test content, rewrites it, and then decides which version is better. It is not fraud. It is a closed circle, where whoever marks the homework also wrote it, and it will reliably discover that text tailored for models is preferred by models.

The strongest study in this field, the 252,000 trials that produced the price result, has exactly this property, and its authors say so. They reread 300 of the scenarios by hand as a check. We cite that study constantly and it has this flaw, which is why we say every time that it has not yet been checked by other researchers.

Ask: who wrote the test material, and who graded it?

3. What is the denominator?

“Share of citations” is usually worked out only across the answers that cited somebody. But most answers cite nobody. Similarweb found that in May 2026 only 7 in every 100 questions asked of ChatGPT in the US produced an answer with any source in it at all. A year earlier it was fewer than 2, so it is growing fast, but it is still 7.

So a headline saying a brand takes 30 citations in 100 may be describing 30 in 100 of almost nothing. Nobody writes that part down, and you will not see it unless you ask.

Ask: 30 in 100 of what, and how many answers cited nobody?

4. Is everyone else doing it too?

Almost every study tests one player changing one thing while the rest of the world holds still. Real markets don’t work that way. One study, C-SEO Bench, which was checked by other researchers before publication, something most work in this field cannot claim, found that most of the techniques on offer “are not only largely ineffective but also frequently have a negative impact on document ranking”, and that they worked less and less as more competitors adopted them. In that market, the authors write, what one player gains another loses.

Which means a tactic that worked in a study run in early 2025 may not work now, for the boring reason that everybody read the same study.

Ask: what happens when your competitors do this too?

5. Is it measuring visibility, or a proxy for visibility?

The survey’s bottom line is worth quoting whole:

“No reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior.”

How often you are cited, how often you are named, how much room you take up next to your competitors: these are all things near the result, not the result. What anyone actually wants is customers, and nobody in this field has yet shown the link between the two.

Ask: what is the outcome being measured, and how far is it from money?

---

Now apply it to us.

Our free report asks ten questions of ChatGPT, twice each. Here is how it scores against its own checklist.

Being found: we measure live answers to real questions, so being found is part of the measurement rather than something we assume. Good.

Judging: a model pulls the names and positions out of each answer, but it never works out a score. The arithmetic is done in code, from stored data. That avoids the worst of the circularity, though the extraction step is still a model reading a model.

Denominator: we report the number of answers that named nobody at all, alongside the ones that named someone. That number is often the most interesting one on the page.

Competition: we can’t control for it. Nobody can. If everyone in your category does this work, the advantage compresses. We would rather say so than sell you permanence.

Proxy distance: we measure whether a machine names you. We do not measure whether that produced a customer, and we will not claim to. That is the honest boundary of what this service can prove, and any vendor telling you otherwise is describing a causal chain the literature has not established.

One last note on the grades we give studies. Most of the strongest results in this field, including several we lean on, were published before other researchers had checked them. We say so every time. When someone shows you a study without saying which category it falls into, that omission is itself information.