Why the GTM check has no AI score
First-hand: Our own check results — the legacy scan's 68-point "AI assessment" moved by eight points between two runs of an unchanged site, and six of its deterministic checks had silently scored every site wrong at some point. The rebuilt check keeps the 32 measured points and drops the opinion.
A score that changes when nothing on your site changed is not a measurement. We ran the old GTMRun scan twice on the same site, one minute apart. The deterministic half scored the same both times. The "AI assessment" half scored 8, then 16. Nothing had been deployed in between. That is why the rebuilt GTM check has no AI score: every row is pass, fail or not measured, and nothing else.
What the moving number cost
The point of a check is to tell you what to fix and, afterwards, whether the fix worked. A row that reads "SPF record authorizes your senders: fail" and then "pass" after you add the record is proof. A total that drifts by eight points on its own buries that proof: you cannot tell your fix from the noise, and you cannot fail a deploy on it in a GitHub Action without failing half the time for no reason.
The second cost was quieter. When a scoring agent failed, the old scan returned a 20-point placeholder that looked exactly like a measurement. Founders saw a number; nobody saw that the number was a shrug.
What "not measured" means
The deterministic checks had their own version of the same mistake, and we found six of them in the old code:
- The e-mail checks shelled out to
dig, which does not exist in the serverless runtime, so every site scored
0/10 on e-mail authentication.
- Hosts that answer every path with the home page were credited with pricing pages that did not exist.
- A domain with a null MX record or a revoked DKIM key was credited with working mail.
- E-mail was graded on the scanned subdomain instead of the zone that actually sends.
- Page weight was measured on a capped download that could never exceed the limit.
- Analytics bundled inside JavaScript was scored "missing".
Each one taught the same lesson: "can't see" is never "no". So the rebuilt check has a third state. When the raw HTML is an application shell, a DNS lookup fails, or a measuring tool does not answer in time, the row says not measured, earns no points, and removes its points from the denominator. The result reads "23 of the measurable 25 points", never "23 of 32" when seven points were unobservable. Every such row links to a guide that says how to verify it yourself.
What replaced the AI half
Judgment did not disappear; it moved to where a person approves it. The check stays deterministic. The drafts — what to fix in which order, a post built on something only your product knows, a reply to a thread where buyers are asking — are cards you approve one by one. A draft can be wrong and cost you nothing; a score that is wrong costs you the one thing a check is for.
Three rows do come from other tools (a rendered-page audit, an image budget, the public link graph), and they are labelled with their source and worth zero points. Third-party measurements are never presented as Google's view; Search Console is the measurement of record.
Try it on your own site
Run the free check at gtmrun.co: no account, about twenty seconds, every finding with a guide. Fix one row, press re-check, and watch it flip. That flip is the whole product.