Budget increases above ~20% per change reset learning ~3x more often than smaller steps

Budget increases above ~20% per change reset learning ~3x more often than smaller steps

Learning reset = a big edit pushes the ad set back into the unstable learning phase.

Scaling-method comparison, 25 accounts:
— +20% every 48-72h: re-entered learning ~12% of the time, CPA stable
— +50% jumps: re-entered learning ~38% of the time ▲, CPA spiked then recovered
— Duplicating to scale lost the original's optimization history — usually started worse, not better
— Horizontal scaling (new audiences) beat vertical jumps for spend over ~$500/day

The algorithm tolerates nudges, not lurches.

Read: scale in ≤20% steps every 2-3 days; big jumps and duplicate-to-scale both restart the learning clock.