A screenshot is not a measurement.

Somebody in the marketing team asks an assistant to recommend firms in your category. Your firm appears. The screenshot goes into a channel, somebody says the AI work is paying off, and everyone moves on.

The following week the answer is different, and nobody checks.

We ran a fixed prompt set weekly for six months, across four assistants and eleven categories, and recorded every firm named. The distribution is the finding: most firms named at all were named rarely, and a small group was named almost every time. The full method and the numbers are in the report.

What the distribution looks like

Being named once is common. Being named consistently is not, and the gap between them is the whole finding.

The largest group by far appeared in only one or two runs out of the whole period. A much smaller group appeared in nearly all of them. Between those two sits everybody else, and the gap between the ends is the whole point.

Every firm in that group could have produced a screenshot. Most of them, asked the following week, would not have appeared.

Most firms that could show you evidence of appearing in AI answers do not hold a position. They were named once.

Three things one check cannot tell you

Which assistant. Most firms named at all were named by exactly one of the four, and only a handful by all of them. So a screenshot from one assistant most likely describes only that assistant.

Which phrasing. We ran variants a reasonable person would call equivalent. The closest kept most of the baseline’s top five; the most divergent kept fewer than half, meaning the majority of named firms differed on a question that meant the same thing.

Which week. One assistant held its top five fairly steady week to week. Another changed most of its list in a typical week. Both were answering the same question.

What a measurement requires

Four things, none expensive.

• A fixed prompt set, written down. Not one prompt — a set covering how a buyer might actually ask. Fixed, because changing the wording changes the answer, and a baseline you cannot repeat exactly is not a baseline.
• More than one assistant. Three or four. Measuring one assistant and generalising is the commonest error, and given how little the four overlap, a large one.
• Repetition over weeks. A few weekly runs across three or four assistants tell you whether you are in the stable set or the incidental one. A single run tells you about a moment.
• Recording what came back, not just whether you appeared. Which firms were named, in what order, and what was said about them. Your own presence is one data point; the shape of the answer is the useful part.

That is an afternoon to set up and twenty minutes a week to run. Any supplier charging for AI visibility work who cannot show you their prompt set has not produced a measurement, and that is the first thing to ask for.

What separates the consistent from the incidental

The consistently-named firms share one thing the incidentally-named do not.

Almost all of the consistently-named were cited in writing they did not control. Almost none of the incidentally-named were. That gap is wider than any other we measured, and it is the one nobody can buy.

Crawlability was near-universal in both groups. So was schema markup. Publishing a blog monthly was, if anything, slightly more common among the firms nobody names. None of the structural characteristics separated the two groups.

The second-largest gap is whether the firm had published original research. That one separates the groups almost as sharply as citation does.

Why this matters commercially

The single screenshot produces two expensive errors.

Stopping too early. A firm sees itself named, concludes the work is succeeding, and stops. Nothing was established and the position was never stable.

Buying the wrong thing. A supplier runs one prompt, screenshots it, presents it as a baseline, and sells six months of work against it. At the end they run it again. Whether the number moved tells you nothing, because the measurement had no stability in either direction.

We sell this work, which is worth stating plainly. It is also why the finding that most changed our own advice is in Report 03: across the firms doing structural visibility work, six months moved the structural characteristics substantially and moved the outcomes very little.

What to do this week

• Run the check properly once. Four assistants, one category, your baseline prompt. Twenty minutes. Write down every firm named, in order.
• Do it again next week. The difference between the two lists tells you more than either list alone.
• Check your robots.txt. One firm in our set was invisible to every assistant for two years because an agency had added a line in 2024 to block scraping. Nobody at the firm knew. Four minutes, and occasionally the entire answer.

The same problem in a different form — measuring the flattering thing rather than the true one — is covered in what a proof of concept actually proves.

The counts, the method and the limits are in What assistants recommend, free to read in full.

Run the visibility assessment · Read Report 03