Three outage days can look like a month of lost visibility

The fixed four-model panel recorded 738 failures in August. The failure rate reached 69.9% on 13 August, 45.6% on 20 August and 20.2% on 28 August. Every other day in that panel completed without a failed request.

Removing the same day numbers from July and August leaves 15,232 August attempts, no failures and only 3 attempts without captured text. The primary content findings keep the same direction in that sensitivity view.

If missing answers are treated as brand absences, an outage becomes a false visibility decline. If failed attempts are silently removed, the report hides where its evidence is weaker. The remedy is to publish collection success and usable-answer yield beside the content measures.

A matched-day sensitivity is useful when an incident is concentrated on known dates. It does not reconstruct the missing answers, but it tests whether the conclusion depends on the affected collection window.