How learning phase thresholds work
When you create an ad set, or make a change big enough to count as a significant edit, Meta’s delivery system starts again from nothing. It does not yet know who converts, on which placement or at what time of day, so it explores. The learning phase is that exploration period, and the threshold is the count of optimisation events the ad set has to record inside a rolling window before delivery is treated as stable.
The event being counted is the one you chose to optimise for, not any conversion. An ad set optimising for purchases needs purchases; the same ad set optimising for add to cart needs add to cart, which happens far more often and is therefore reached sooner. Until the threshold is met the Delivery column reads Learning, and where the ad set plainly cannot get there it reads Learning limited instead.
Why learning phase thresholds matter
They shape the structure of an account more than any other single rule. Events are the currency, and splitting a modest budget across many ad sets divides that currency until none of them accumulates enough. This is a real constraint in Nepal and in other smaller markets, where audiences and budgets are smaller and an ad set optimising for purchases can sit in Learning limited indefinitely.
They also explain volatility. Cost per result swings while the system is still exploring, so judging a campaign on its opening days is judging the exploration rather than the outcome.
Where learning phase thresholds go wrong
The threshold gets treated as a promise. It is not one. It is Meta’s own guidance for the point at which delivery stops fluctuating, published as a rule of thumb; it does not say your cost per result will improve once you cross it, and it is not a benchmark to hold an account to. Plenty of ad sets exit learning and still lose money, and some ad sets stuck in Learning limited deliver acceptable leads month after month.
The second mistake is restarting the count without meaning to. Changing the budget substantially, swapping the audience, editing the optimisation event or replacing the creative can all reset learning, so an account that is edited every day never leaves it.
What to do about it
Consolidate first. Fewer ad sets, each carrying more of the budget, concentrate events where the system can use them, and campaign budget optimisation lets the platform move money to whichever one is working. Then pick an optimisation event that genuinely occurs often enough to be counted; for a low-volume business that usually means a lead or an add to cart rather than a purchase.
After that, leave it alone. Set a review point far enough out that the exploration has finished, note what the Delivery column says when you reach it, and only then change something. If you need the current threshold figure, read it in Meta’s own delivery documentation rather than in a blog post, because the guidance is revised. Structure is decided at setup, and it is the part of Meta ads management that is hardest to correct later.
