Definition
What is the Meta learning phase?
The learning phase is the period during which Meta's delivery system explores different audiences, placements and times of day to find who responds to an ad. Meta documents its exit at roughly 50 results in the week following the last significant edit to an ad set. While it lasts, costs swing widely: that is the system working normally, not a sign of poor performance.
How to recognize it
In Ads Manager, the delivery column shows “Learning” for an ad set that has not yet reached the threshold. The “Learning limited” status flags an ad set that is not producing enough results to get out. That second status points to a volume problem: redoing the creative does not fix it.
What restarts it
Meta states that certain edits restart the learning phase from zero:
- a change of targeting;
- a change of the optimization event (the action Meta tries to get, such as a purchase or a sign-up);
- a change of bid strategy;
- frequent budget changes.
A team that tweaks every other day keeps its ad sets in permanent learning. It then reads unstable costs as if they reflected the market.
Why the number of ad sets matters
Each ad set learns separately. Meta writes that the more ad sets an account creates, the less information its system gathers on each one, and recommends combining similar ad sets. The same budget spread over four ad sets has to produce the exit threshold four times.
Minimum weekly budget for one ad set = 50 × target cost per result
Run the math before launching. If it exceeds the budget, optimize for a more frequent event (a sign-up rather than a purchase) or test fewer ad sets at a time.
An online store targets sales at $25 each. For a single ad set to exit learning, it needs about 50 sales in a week, which means roughly $1,250 of weekly budget.
If it spreads that same amount over three ad sets, each will produce about 17 sales a week. None will reach the threshold, and the account is likely to stay in learning however long the test runs.
We count a test in results, not in days. An ad set producing seven sales a week can stay in learning for two weeks or six months: below the threshold, no duration lets you draw a conclusion. The budget for a test is therefore set from the cost per sale, before the first dollar is spent.
We also set in advance which edits are allowed and who can make them. Well-meant tweaks made during learning are the leading cause of tests that never reach a conclusion.
Not to be confused with
- Google Ads learning period
- Google uses a similar notion for its automated bidding strategies. Its documentation calls for at least two weeks without edits for the initial period, up to four weeks on the display network. The exit threshold is not Meta's.
- Useful conversion threshold
- The useful conversion threshold is a monthly volume below which an automated bidding strategy cannot hold its target reliably. The learning phase is a period; the threshold is a number of conversions.
Related concepts
- Meta advertising
- Useful conversion threshold
- Floor budget of a market test
- CPA
- Audience fragmentation
- Freeze date
- Andromeda
Further reading
- How long does it take to test a Facebook ad?
- Meta advertising: how to diagnose a Facebook and Instagram account
- Facebook targeting: Meta itself documents why fine-grained targeting hurts
- Black Friday and the holidays: the decisions are made before, not during
- New market: the campaign that learns comes before the campaign that sells
Related services
Frequently asked questions
How long does the learning phase last?
It is not counted in days. Meta documents the exit at roughly 50 results in the week following the last significant edit: an ad set that reaches that volume gets out in a week, and one that does not can stay in it indefinitely.
My costs are very high in the first few days: is that normal?
Yes. During learning, the system tries several audiences and placements: some attempts cost a lot while others cost little. Judging performance at that point can lead you to cut an ad that was about to settle.
