Measurement

We Checked 32 Companies for AI Visibility. 27 Were Never Named Once.

44 buying questions, 111 runs, one model. Here is the whole dataset — including the reason you should not read the headline number as a rate for small companies generally.

By Thicket AI Team · August 3, 2026 · 7 min read

Across 44 category questions and 111 usable runs against one model, 27 of 32 companies were never named once. That number is real and we are publishing the whole of it. It is also not a population rate, because we chose these companies partly for their likely absence — and saying so is more useful than the headline.

The numbers

Measure	Value
Companies checked	32
Distinct buying questions	44
Usable model runs	111
Companies named at least once	5
Companies never named	27

One of the five named is Notion, which we run as a control precisely because a model that failed to name Notion would be telling us our checker was broken rather than telling us anything about Notion. Excluding it, four of thirty-one small companies were named.

Why you should distrust the headline

We did not sample these companies at random. We check a company when we are considering contacting it, and the reason to contact it is that we expect it to be absent. The sample is selected on something close to the outcome being measured. If you took the 84% figure and reported it as “AI ignores 84% of small companies,” you would be repeating a number that our method manufactured.

We are stating this prominently because it is the single most likely way this page gets misused, including by us. A dataset that only supports a narrow claim should be published with the narrow claim attached.

What survives the bias

Two things, because neither depends on how the companies were selected.

1. Absence is consistent, not intermittent. Where a company was absent, it was almost always absent across every differently worded question we tried — seven questions and sixteen runs for one company, four questions and twelve runs for another, zero mentions in both cases. This matters practically: it means a single check is reasonably informative about a company’s standing, even though a single run is not. Absence looks structural rather than random.

2. Comparison pages did not rescue anyone. This is the finding we did not expect and did not select for. Three of the companies we examined most closely already had extensive comparison pages — one had nine, each targeting a named competitor, server-rendered and indexed. All three were absent anyway, including from the exact query their page was written to answer. We wrote about the first of those cases in detail in we told a startup to build comparison pages, and it already had them.

Selection bias explains why so many companies in our set are absent. It does not explain why the ones who did the recommended on-page work are absent too. Those companies were not selected for having done the work — we found that out afterwards, in one case because the founder corrected us.

The working explanation

A model answering from training knowledge reflects what the web said about a product, not what the product said about itself. Every vendor publishes a page claiming to be the best alternative to its biggest competitor. That claim carries almost no discriminating information precisely because everyone makes it. What seems to separate the named from the unnamed is whether somebody else said it.

We hold that as the best available explanation, not a demonstrated mechanism. We cannot see inside the model, and neither can anyone selling you a service based on claiming they can.

How to run this yourself

Ask the question your buyer asks, never your brand name. A model will discuss a brand named in the prompt while never surfacing it unprompted.
Run each question at least three times and record the count.
Record who was named. The competitor list tells you which material the model learned from, which is more actionable than your own absence.
Treat a failed request as an error, never an absence. This is the one that bites: we ran a metric for three months that reported 0% off 280 consecutive failed API calls, because a failure and a zero were stored identically.
Check what you already have before acting on any recommendation, including ours.

Method

One model (a Gemini Flash release), answering from training knowledge, no live web search, between 24 July and 3 August 2026. Each question asks for specific product recommendations with a one-line reason. Brand matching is done on name and domain with word boundaries, so a short brand name cannot match inside a longer word. Runs that returned no usable response are recorded as errors and excluded from the denominator rather than counted as absences. Nothing here generalises to assistants we did not test, and model behaviour changes without notice.

Frequently asked questions

How often do AI assistants name small companies in category answers?

In our own records, 27 of 32 companies were never named once across 44 buying questions and 111 runs. But that number should not be read as a rate for small companies generally, and we would be misleading you if we presented it that way. We selected these companies partly because we expected them to be absent — absence is what makes our outreach relevant — so the sample is biased toward the result it produced. What the data supports is narrower: when a small company is absent, it tends to be absent consistently rather than intermittently, across differently worded questions.

Does having comparison pages get you named by AI?

Not on its own, and this is the pattern that survived our selection bias. Three of the companies we examined most closely had extensive comparison pages already live — in one case nine of them, targeting each major competitor by name, server-rendered and indexed. All three were still absent from the answers those pages targeted, including the exact query the page was written for. On-page work appears to be necessary but not sufficient. It makes you eligible to be named rather than causing it.

Why would a model name a competitor instead of a company with a better page?

Because a model answering from training knowledge is reflecting what other people wrote about a product, not what the product wrote about itself. Every vendor in a category publishes a page claiming to be the best alternative to its largest competitor, so that claim carries almost no discriminating information. Third-party mentions — someone else naming you, in a context the model ingested — appear to do the work instead. We hold that as the best available explanation rather than a demonstrated mechanism, because we cannot see inside the model.

How many times should you run a query before trusting the result?

At least three, and you should report the count rather than a verdict. Model output varies between samples, so a single run can miss a company by chance and produce a confident false absence. We record runs attempted, runs that returned a usable answer, and runs that named the company, specifically so that a failed request can never be silently recorded as an absence. That distinction is not academic: we previously ran a metric of our own that reported 0% for three months because 280 consecutive failed API calls were stored identically to genuine zeros.

Is this true of ChatGPT and Claude as well?

We do not know, and nothing here should be read as a claim about assistants we did not test. Every measurement in this piece comes from one model answering from training knowledge, without live web search, in July and August 2026. A different assistant will have ingested different material, and the same assistant with web grounding enabled behaves differently again. Anyone generalising a single-model result to artificial intelligence as a whole is overreaching, and that includes us.