Many Meta ad managers experience the same pattern. In the first week, results fluctuate, the cost per result feels high, and then the ads are edited to improve quickly. A few days later, the pattern repeats from the beginning. What usually happens is not that the ads worsen, but rather that the learning phase count restarts without notice.
The learning phase is the period when the Meta display system is still learning the best way to show a set of ads. During this period, the Delivery column in Ads Manager reads "Learning". Meta states that performance during this phase is indeed less stable and the cost per result tends to be higher, so the figures in the first week do not reflect long-term performance.
What is counted is not clicks, but optimization events
The most commonly misunderstood point: what the system counts are optimization events, which are the results you choose to optimize at the ad set level, not the campaign objective. The two can differ. You can choose sales as the campaign objective but optimize the ad set for link clicks. That event is what the system bids on in the auction, and its counting follows the conversion window you set when creating the ad set.
The threshold is around 50 events in one week
An ad set exits the learning phase once its display stabilizes, which usually occurs after about 50 results in one week since the last significant edit. Meta places the number 50 as a recommendation, not a hard rule, and mentions that some ad sets stabilize faster.
There are exceptions with specific numbers. For Shops ads, the learning phase is only considered complete after a minimum of 17 purchases through your site and 5 through Meta.
List of edits that restart the count to zero
Not all edits restart the learning phase, only those classified as significant. Meta explicitly lists them:
- any changes to targeting
- any changes to ad creative
- any changes to optimization events
- adding new ads to the ad set
- pausing the ad set for seven days or more, and the count restarts once the ad set is reactivated
- changing the bidding strategy
Three other factors can be significant or not, depending on the magnitude of the change: ad set spending limit, cost control or ROAS target, and budget amount. The examples provided by Meta clarify the limits: increasing the budget from 100 to 101 dollars is unlikely to trigger a restart, while increasing from 100 to 1,000 dollars could.
In campaigns using Advantage+ campaign budget, changing the bidding strategy or campaign budget can cause multiple ad sets to restart the learning phase simultaneously. Conversely, automatic budget distribution among ad sets does not trigger it, and adding a new ad set does not affect other ad sets in the same campaign.
When the status reads Learning limited
If an ad set is expected not to gather around 50 optimization events in a week after the last edit, its status changes to "Learning limited". Meta emphasizes that this status is not a penalty, but a sign that the budget has not been used effectively.
Causes mentioned by Meta include: audience too small, low budget, bids or cost controls too low, high auction overlap, optimization events occurring too infrequently, and too many ads running simultaneously. Solutions follow the causes: merging ad sets and campaigns, expanding the audience, increasing the budget, raising bids or cost controls, or selecting optimization events that occur more frequently, such as switching from purchases to add-to-cart events. Once enough events accumulate since the last edit, the status returns to Active.
Two columns that should be displayed
All of the above can only be monitored if the columns are displayed. In Ads Manager, add the "Last significant edit" column to see when you last triggered a restart, and the Results column to see the number of results since then. Optimization events can be added via Columns, then Customize columns, and check Optimization events.
With these two columns, the question "why are the results unstable" turns into a question that can be answered by numbers: how many events have accumulated since the last edit, and whether the next edit is truly necessary.