Four steps, then it gets on with it.
1. Connect an account
You send your client a link, or connect your own account. They approve through Meta. No Business Manager navigation, no shared logins, no password ever reaches us.
2. Set a target CPA
The one number that matters, and the one most tools quietly default to the account average — which flags roughly half of any account by construction and means nothing. Set yours from what a conversion is actually worth.
3. Watch it in dry run
It reads the last 30 days and tells you what it would do, with the evidence. Nothing changes. Most people leave it here for a week, and that is the right instinct.
4. Let it act
Switch on pausing when you trust it. Budget increases stay behind their own switch, +25% per step, 48 hours apart, hard-capped. You can turn any of it off in one click.
The rules, in full
No black box. These are the thresholds, and they are the same ones the engine was backtested against — imported from the backtest, not re-typed, so the two cannot drift.
| Rule | Fires when |
|---|---|
| Dead ad | £500 spent with zero conversions |
| Over target | £300 spent, 10+ conversions, and CPA is 30% worse than target |
| Decaying winner | 7-day CPA 50% worse than target, even while lifetime CPA still looks fine |
| Fatigue | Frequency above 4.0 alongside a rising CPA |
| Scale up | 15+ conversions at 20% better than target — then +25%, once per 48h |
The evidence bars are deliberately conservative. A real buyer would act sooner; we would rather be late and right than early and wrong with someone else's money.
What we measured before building it
We replayed these rules over £437,790 of real spend and 19,021 conversions, walking forward day by day using only the data that existed on each day. We report spend-after-signal: money that went to an ad already provably past target on the evidence available at the time. We do not quote a "leads we would have gained" figure, because that requires assuming how the winner scales, and nobody can measure that. Every vendor quotes it anyway.