Experiment monitoring and automation

You are told, rather than finding out from the numbers weeks later. The two failures that are invisible from inside a results screen are a split that stopped matching what you configured, and an experiment that quietly stopped recording anything, usually because a deploy took the snippet or a targeting rule stopped matching. Both raise an alarm to email and Slack by default, both mark every result on that experiment as untrustworthy, and both stop the platform acting on it. The work waiting for you is a ranked queue rather than a mess, and it is worked out from what is currently true, so nothing disappears because somebody cleared a notification.

Nothing, until you turn it on, and then less than you might expect. Automation and automatic rollout of a winner are separate switches and both start off, because auto-stopping the wrong experiment costs the traffic it already spent while auto-rolling-out a loser changes what every visitor sees. Before it acts it has to clear three trust gates and a minimum runtime. And with automation off it still evaluates everything and records what it would have done, using the same code that would have done it, so you can watch it be right for a month before you let it act.

Members are read-only across the entire platform: they can see every experiment and every result and change none of it. That rule is enforced once, centrally, rather than screen by screen, which matters because a permission model applied per screen is one where the next screen built is quietly the exception. Starting, stopping and rolling out stay with admins and owners, and every change that matters is recorded against the experiment with a name and a time on it.

That is the shape the monitoring is built around. Everything unresolved across every account you can reach is rolled up in one view, grouped by client and by website, ordered so that anything untrustworthy comes before anything merely waiting on a decision. The scope comes from who you signed in as rather than from an identifier in the address, so there is no account id in the URL to check and none to tamper with. An agency sees its roster; a single-brand customer sees their own account in the same screen.

Alerts go to email, to Slack, or to neither, set per event type and per person rather than once for the whole account, and Slack messages link back to the exact screen the event refers to. On the measurement side each website is set to GA4, Matomo or Pertento native tracking, which means the conversions we count can be the same events your analytics already counts rather than a second definition of a purchase that somebody has to reconcile.