Model Visibility

People increasingly ask an assistant about an API instead of reading its documentation. So we ask three of them, on a schedule, about companies in this catalog — and check the answers against the artifacts those companies actually publish.

Last run 2026-09-08 · wave 1 · 65 providers · claude-opus-5 · gpt-5.5-2026-04-23 · gemini-flash-latest

What we measure

Every question is asked twice, and the two answers mean different things. With search enabled we learn what a model can find. With tools off we learn what it already knows. Those move on completely different clocks, so they are never pooled.

How it works

We ask closed questions, so an answer can be checked by a rule rather than by another model’s opinion:

• Does <company> publish a public OpenAPI specification?
• Does <company> operate an MCP server for its API?
• Does <company> offer a free tier or trial for its API?

Each answer is compared against that company’s own published artifacts, as recorded on its provider page. Where we do not hold the fact, nothing is scored — a gap in our profile is our backlog, not the model’s error. Saying “unknown” is counted separately and is not treated as a wrong answer: a model that declines to guess about a small company is behaving correctly, and a metric that punished that equally would be recommending confabulation.

The providers in this wave

Each row links to the provider page holding the artifacts every answer was checked against, so any claim here can be re-checked rather than taken on trust. A low score is a finding about the assistant, not about the company.

Reading this honestly

• A wave is a sample. One provider changing its mind is a few percentage points. Read “not observed in this sample”, never “the model does not know”.
• The three assistants are never averaged. They disagree about the same company, and the disagreement is the point — a company fixing one of them fixes it in a different place than the others.
• A model version change breaks the series. Two of the three are pinned to a dated build. One is a moving alias, so its line before and after a version change is two series, not one trend.
• We are one of the sources being measured. “Cited us” and “used our vocabulary” are numbers about our own reach, and we report them here rather than only in private because they are the least flattering numbers on the page.

The data behind this page is public. model-visibility.json carries every claim, what each assistant said, what we hold, the exact model ids and the capture timestamp — so nothing here has to be taken on trust.

If a result here is wrong about your company, that is worth telling us — the fastest fix is usually the artifact the model could not find. Claim or correct your listing →