The on-call engineer who silences a page without reading it isn't negligent — they've correctly learned the alert carries no information. Alert fatigue is a rational response to a broken signal, not a discipline problem.
If your team has stopped treating pages as urgent, start by measuring the actionability rate: what proportion of pages in the last 90 days resulted in a human doing something? Below roughly 30%, the on-call rotation is being trained to ignore you. That number matters more than MTTD or MTTR, which are lagging indicators of a system that may already have taught people the wrong lesson.
Three kinds of alert that get conflated
Something is broken now. Something will break soon. Something is interesting. Only the first should page a human at 3am. The other two belong on a dashboard and in a ticket queue. When those three share the same channel, every alert becomes optional, including the ones that aren't.
Every alert needs a named owner and a runbook link, or it gets deleted. No exceptions, including for alerts somebody is fond of. Ownership without a runbook still leaves the next person guessing; a runbook without an owner becomes folklore.
Delete first, instrument second
Teams almost always add coverage before cutting noise, which makes the signal worse. The counter-intuitive move is to delete aggressively first, then instrument the gaps that remain with clear thresholds and owners. Coverage you cannot act on is not coverage — it is theatre.
MTTD and MTTR remain useful once the channel is trustworthy. Until then, the leading indicator is simpler: whether your on-call engineers read the alert before acting on it.