Platforms need a threshold volume of optimisation events — commonly around fifty per ad set per week on Meta — before delivery stabilises. Below that, the system is effectively guessing, and results swing widely from day to day for reasons that have nothing to do with your creative or your audience.
Any significant edit restarts it. Budget changes above a modest threshold, new creative, changed audiences, changed optimisation events and changed bid strategy all reset the clock. This is the mechanism behind most self-inflicted account damage: a team sees poor early numbers, edits, resets learning, sees poor numbers again, and edits once more.
A campaign that never accumulates enough events sits permanently in a limited state, delivering unpredictably and expensively. That is a structural problem — too many ad sets splitting too few conversions, or an optimisation event that fires too rarely — and no amount of patience fixes it.
Why Learning Phase matters
It sets a hard floor on how fast you can iterate, and it is the reason most accounts underperform their potential. Teams that test weekly are resetting learning weekly and never seeing a stabilised result from anything — the account is permanently measuring noise and acting on it.
Consolidation as the fix
An account runs nine ad sets, each generating around fifteen conversions a week — every one of them stuck below the stability threshold. Consolidating to three ad sets puts each at roughly forty-five weekly conversions, close enough to exit learning. Nothing about the creative, the audience or the budget changed; only the structure did. That change alone commonly produces a double-digit percentage improvement in cost per acquisition, purely from delivery stabilising.
Common mistakes
Editing during learning because results look bad
Early performance is supposed to look bad — that is what the phase is. Editing in week one restarts it and creates a loop the account may never escape.
Too many ad sets for the conversion volume
Splitting audiences feels like control, but it divides the events each ad set needs to stabilise. Fewer, larger ad sets almost always beat more, smaller ones at modest volume.
Optimising for an event that rarely fires
Choosing a deep event like a qualified lead or a subscription when it happens ten times a week guarantees permanent learning limitation. Optimise to a higher-frequency event that still correlates with revenue until volume supports the deeper one.
Assuming learning is over because the label cleared
The status indicator is a threshold, not a guarantee of stability. Performance often keeps drifting for a week or more after the label disappears, so judging a test the moment it clears is still premature.
Where we work on this