How the learning phase works
When an ad set is created, or changed in a way Meta counts as significant, delivery starts again from scratch. During that spell the system is exploring: trying placements, moments and audience pockets to work out who responds to this particular combination. Costs swing, results are uneven, and the numbers you are looking at describe an experiment rather than a campaign.
The phase ends when the ad set has accumulated enough recent optimisation events — conversions of whatever type you told it to optimise for — inside a rolling window of roughly a week. If it cannot gather them, delivery is labelled learning limited and stays in the unsettled state indefinitely. What resets the clock is a substantial budget change, new targeting, a different optimisation event, a change of bid strategy or new creative.
Why the learning phase matters
It explains why the first days of an ad set are not the results of the ad set. The most expensive habit in Meta advertising is judging a campaign early, editing because the cost looks wrong, restarting learning, and repeating that cycle until the account has never once been allowed to settle. Spend goes into exploration over and over, and the settled performance nobody waited for never arrives.
Learning limited matters differently. It is a structural signal, not a temporary one: it says the ad set cannot get enough conversions to be predictable, usually because the budget is small relative to the cost of a conversion, the audience is too narrow, or the chosen event is too rare. That combination is common on modest Nepali budgets optimising for purchases.
Where it goes wrong
Daily tinkering is the obvious one. Less obvious is fragmentation: an account split into many ad sets, each carrying a slice of the same budget, where none of them individually gathers enough events to leave learning. The structure looks organised and performs badly.
The third mistake is optimising for the rarest event in the funnel because it is the one that matters commercially. A high-value purchase that happens occasionally will not train anything. The fourth is the opposite belief — that leaving learning is a guarantee. It only means delivery has stabilised, not that the results are good.
What to do about it
Decide before launch how long the ad set will run untouched, and hold to it. Batch changes into a single deliberate edit rather than adjusting something every morning, and where you want to test a genuinely different idea, duplicate the ad set instead of editing it, accepting that the duplicate begins its own learning.
If learning limited persists, treat it as a structural fix rather than a bidding one. Consolidate ad sets so the budget concentrates, widen a narrow audience, or move the optimisation event to something that happens often enough to train on while checking that the deeper value still follows. Steady, patient management of a Meta ads account is mostly the discipline of leaving things alone long enough to know whether they work.