Averaging 44 checks is how a noindex ships: the case against audit health scores

Do the arithmetic on the composite before trusting it as a gate. One check out of 44 is worth 2.3 points of an unweighted average ^[1]. Set the release threshold at 95 percent and you have pre-authorised two failures, whichever two they turn out to be, because 42 of 44 is 95.5 percent ^[2]. The average has no way to know that one of those two is a canonical pointing at a redirect or an accidental `noindex` on a key template, which is precisely the class of defect Edikka lists as release-stopping ^[2].

The reason is structural. Severity is an ordinal label, not a quantity, and a mean requires commensurable units. Edikka's protocol keeps the two apart: each applicable check carries a status from four options and a severity from four more ^[5], and the decision reads that pair rather than a total ^[6]. A score cannot represent "this single item is disqualifying" without weighting so extreme that the number stops being an average and becomes the rule in disguise.

The awkward half of the rule is the treatment of absence. A Blocking check that was never run produces NO-GO on the same footing as one that was run and failed ^[6], because Not tested is not Compliant ^[8]. A percentage does the opposite by default: unrun checks either vanish from the denominator or get quietly scored as passes, and the release inherits the gap.

That distinction matters most where audits reach past what they can see. A public audit can establish that a page declares a coherent canonical; it cannot claim Google selected that canonical without Search Console evidence ^[10]. The same discipline applies further down: a 200 response proves the server returned a successful representation, not discovery, indexation, or selection for a query, and a sitemap declares candidate URLs rather than certifying indexation ^[11]. Edikka also warns against trusting a HEAD request alone, since applications, CDNs, and firewalls can treat HEAD and GET differently ^[12], and against the happy-path crawl: the expected outcome across a canonical URL, a host variant, an old redirect, a removed resource, and a URL that never existed is the intended status for each case, not 200 everywhere ^[13].

Mechanism beats intent here, and the `noindex` case shows why a scored crawl misses it. Google's robots directives documentation, as cited in the piece, notes that crawlers must be permitted to reach a page in order to read its `noindex` rule, and that the Robots Exclusion Protocol defines crawling rules rather than a security boundary ^[14]. A robots.txt block plus a `noindex` tag can each look defensible in isolation while combining into a page that is neither crawled nor de-indexed on purpose. Catching that requires comparing the raw HTML response, the rendered DOM, and URL Inspection for the same template ^[15], which is evidence collection, not a metric.

Two caveats belong on the record. Edikka states plainly that this is its own governance rule and not an industry standard, that teams may set stricter thresholds, and that the rule has to be written before the result is known ^[7]. And the protocol declines to predict rankings, traffic, rich results, or AI citations ^[9], which removes the main thing a health percentage is usually asked to imply.