Your vibe-coded app is up.

An uptime service pings your site every few minutes and alerts you when it stops responding. That is a real failure mode and worth covering. It is also not what happened here. The site was up for all eight weeks. Pages rendered, forms accepted input, and the endpoints returned ordinary-looking success responses. A prober would have recorded perfect availability while the business quietly took no bookings.

The distinction worth carrying away is a single question. Uptime monitoring asks "is it responding?" Outcome monitoring asks "did the thing that should have happened, happen?" Only the second question would have found any of this. So we bought no uptime service. What we built instead queries our own records and compares them against what ought to be true.

List every pipeline in your app. A pipeline is anything where one thing happening is supposed to cause another thing to happen: a form submission that triggers an email, a signup that syncs to a mailing list, a nightly job that produces a report. For each one, write a single sentence in this shape.

When X happens, Y must exist within Z.

That sentence is the check. Once it is written, turning it into a database query is mechanical, and it is a task an AI coding assistant does well because you have removed every ambiguity. Here are four of ours, in the form we wrote them before any code existed.

When someone registers for an event,
  a confirmation email record must exist within 15 minutes.

When someone ticks the consent box,
  they must appear in the mailing list within 1 hour.

When a session is less than 23 hours away,
  every registrant must have a reminder logged.

When a scheduled job is due,
  a successful run must exist within 25 hours.

Notice what the exercise forces. You cannot write the sentence without deciding what "done" looks like and where the evidence of it lives. Half of our failures existed because nobody had ever articulated the second half of the sentence, and a consequence nobody named is a consequence nobody checks.

If you do nothing else from this guide, do this. Ten sentences on a page will tell you more about where your app is quietly broken than any dashboard you can buy.

The temptation is to pick round numbers by instinct. Don't. For each surface, query the gaps between records across the last four months, find the longest gap that fell in a stretch you know was healthy, and set the alert comfortably beyond it. We re-derived every one of ours this way rather than guessing, and several came out nowhere near where instinct had put them.

The shape of the answer is the transferable part. Whatever your volumes, it tends to land like this.

The rule underneath it is roughly one and a half times the longest healthy gap, rounded up to a number you would actually act on. The bottom row is the one people get wrong. On a surface that only sees a handful of events a month, a fortnight of silence genuinely means nothing, and alarming on it would train you to stop reading the email entirely. Report it as context instead, where a human can notice a pattern over time. That is the right home for a weak signal.

If you have less than four months of history, use what you have and revisit it. The threshold is not sacred. What matters is that it came from evidence, because a threshold you guessed is a threshold you will not trust at 7am when it turns red.