Marketing alerts you set by asking: CPA, ROAS and tracking watchdogs

Written by Uptimal, the team that builds and operates Upbuild and runs marketing analytics for multinational clients across the globe. Everything below reflects how the product actually works, including the parts that are deliberately boring.

Nobody reopens the dashboard at 22:40, which is exactly when the CPA doubles. Alerting is the unglamorous half of reporting, and it has always required either an engineer with a cron job or faith in six different platforms’ native rule tabs. The assistant-first version: say the guard out loud, once.

From sentence to standing watchdog

"Alert me if blended CPA rises more than 15% week over week." In Upbuild Enterprise that sentence becomes a stored guard: a definition of the metric (blended, meaning across the platforms you name), the comparison window, the threshold, the cadence it is evaluated on, and where the message lands: Slack, Teams or email. Every morning the condition is computed from live account data through the read-only connector, and you hear about it only when it trips.

The interesting alerts are the cross-platform ones no native rule can express. Blended CPA needs Google and Meta in one fraction. "Search terms spending with zero conversions for a third straight month" is a waste guard across time, not a threshold. "The checkout Floodlight recorded nothing for 30 days" is a tracking alarm, not a media one. One read-only surface across ten platforms is what makes those sentences computable at all.

Anatomy of an alert worth waking for

A good alert carries its own evidence. The message that lands in Slack states the number, the baseline it broke, and the top contributors, because an alert that says "something moved" just schedules the investigation it should have started. And it has to run on infrastructure that assumes failure: in Upbuild the connection layer is probed every few minutes, evaluations run on their schedule with provenance, and an unreachable platform is itself a condition you hear about rather than a silent gap in coverage.

Tuning without the boy-who-cried-wolf phase

Thresholds are conversations here, which changes the tuning economics. "Too noisy, make it 20% and exclude brand campaigns" is a sentence, not a ticket, and the change is a drafted, tested, published revision like every other definition in the system, with history. Teams end up with a small portfolio: one budget guard, one efficiency guard, one waste sweep, one tracking watchdog. Four sentences, maintained by saying so.

Alerts complete the loop

Dashboards answer the questions you knew to ask; conversations answer the ones you did not; alerts ask on your behalf while you sleep. On shared rails (same connector, same versioned definitions, same delivery into the tools you live in) the three stop being separate products. That is the actual promise of the AI analytics stack: coverage without headcount.

Frequently asked questions