Discussion about this post

User's avatar
Andrew Moullin's avatar

Cutting on the right number rather than simply cutting fast is the distinction most accounts miss. The backtest is the part I'd make everyone do, since checking whether your own winners would have survived your rules is the quickest way to find out if the rules are any good. Hopkins tested ad against ad for much the same reason. An opinion about an ad only counts once the data has had a fair chance to disagree with it.

Pedro Alves - PA Analytics's avatar

The spend-scaled cut line is the right frame, and the scatter plot makes it operational. One thing worth adding: CPA on the Y-axis hides the ads that spend slowly but convert later, so the log-scale spend axis helps but the cut still punishes an ad for a timing gap rather than a quality gap. With BFCM five weeks out I'd rather keep a slow starter that clears target on assisted conversions than cut it and re-learn the lesson in November.

The other trap is cutting on platform CPA when the store's own numbers disagree. If the cut decision comes from a number that can't be reconciled with real orders, you're pruning on noise.

1 more comment...

Ready for more?