Why there is no score

No published research establishes those weights. We chose them. They were the shape of a measurement without being one: a page could score 82 and never be cited, and a page could score 41 and be quoted in every answer. The number did not predict the thing it appeared to measure, and it was not reproducible — change the weights and you change the verdict without changing the page.

The failure was not that the arithmetic was wrong. It was that the output was formatted like a fact. A percentage has a denominator and a method behind it, and readers reasonably assume both exist. Ours did not, and the honesty of every other number on the page is what made that one persuasive.

An observation, not a grade. The readiness review reports each signal as present, absent, or unknown, and says plainly when a missing one is a legitimate choice rather than a gap. Nothing is added up. The tool that replaced the score lives at lib/tools/inspect.ts and its own comment records why it changed.

An answer, not a rating. When a brand is invisible in AI answers, the useful question is why. The diagnosis is ordered on purpose: a crawler that cannot read the site is the answer, and everything below it is marked deferred with “nothing else matters until that is fixed”.

“Not measured” instead of zero. If a vendor did not return a value, a request timed out, or a check could not be evaluated from outside, the report says so and keeps that state distinct from a failure.

A score with its denominator, or no score. The checklist score that remains is our own count of the checks that applied, published beside the denominator, null when nothing scorable applied, and never presented as a Google signal.

One percentage in the whole product. Uptime. It has a numerator and a denominator that are both rows in a database — an up count over a total count — which is the only reason it is allowed.

Scores sell. A number that goes up is easier to show a client than a list of findings, and every competitor ships one. Removing ours made the product harder to explain and less impressive in a screenshot. We did it anyway, because a number nobody can reproduce is a liability the moment a reader checks it — and the readers who check are the ones worth having.

It also cost us a feature we could have kept alongside the honest parts. The temptation was to publish it “for guidance”. That is the version that does the most damage: authoritative enough to act on, hedged enough to disown.

The rule is easy to state and easy to check: no invented numbers, ever. Every check definition is published at /methodology with what it measures and what it cannot, and as JSON so it can be diffed. The limits page is the other half of the same promise. And if you find a number on this site that we cannot point at a source for, that is the most serious bug we can receive — report it and it goes on the public record with what we did about it.

What crawlwise cannot tell you · Check methodology · Feedback record