What we publish, and why

Every visibility score you have been shown is a number without a denominator. Someone asked a model some questions, some number of times, and reported a figure. You were not told which questions, how many times, or what came back. You were asked to trust it.

We think that is the whole problem with this category, so this post is the denominator.

The prompt set is public

We run 128 prompts per scan. The set is published, and you can edit it.

That last part matters more than the first. A published prompt set you cannot change is still our opinion of how your buyers phrase things, and we are frequently wrong about that. A freight forwarder we scanned early on cared about one phrase we had not thought to include — and once it was in the set, their score moved eleven points, because the gap had been there all along and we had not been asking.

Every prompt runs sixty times

Ask a model the same question twice and you get two different answers. This is not a defect you can engineer around; it is what the thing is.

So a single run is an anecdote. We run each prompt sixty times and report how often you were named, out of sixty, along with the spread. A brand named in 37 of 60 runs and a brand named in 58 of 60 do not have "similar visibility" — one of them is a coin flip and the other is a default, and a single integer hides that difference completely.

When we quote a score, the sample size and the margin travel with it. If you ever see one of our numbers without them, that is a bug.

The raw answers are kept

For every run we store the verbatim response, the timestamp and the model version. You can read the actual paragraph that did or did not name you.

This is the part that makes us most uncomfortable to publish, and it is the reason we do. If our scoring is wrong, the evidence to prove it is in your account, exportable as JSON or CSV. We would rather hand you what you need to win that argument than win one you cannot check.

What we deliberately do not do

We do not currently measure citation frequency on your behalf across every model in real time, because running 128 prompts sixty times across six engines costs real money per scan. The free scan covers the technical half — whether a model can read, parse and attribute your site at all. Where the number would be invented, the report says so rather than filling the space.

We also do not claim causation we cannot show. If your score moves after we ship a change, we tell you what we shipped and what moved. Sometimes those are not related, and models drift underneath everyone. Saying so costs us a nice-looking chart and buys the only thing that matters here, which is that the next number we show you is believable.

The uncomfortable part

Twenty-six of our 36 technical checks deliberately reuse another tool's check ids, grading rules and point scale. You can open their report for the same host and read the two side by side. Across five reference domains the two engines agree on 88% of checks.

We did that on purpose, and it is not generosity. A score that only exists inside our product is unfalsifiable by construction. Making ours directly comparable to somebody else's is the fastest way to prove we are not quietly grading on a curve — and if we ever disagree with them badly on your site, you will see it, and you should ask us why.