What happens during learning
At launch, the algorithm does not yet know which people, placements, and moments produce conversions. It explores: it tests varied bids and audiences, some of which yield nothing. Results are more erratic, and often more expensive, than they will be later.
The algorithm learns from optimization events: the conversions you asked it to look for. The more of them there are, the faster it learns. A campaign optimized for a rare event, such as a €2,000 sale or a signed lead, learns more slowly than one optimized for add-to-cart.
During this period, the numbers are not representative. Cutting or correcting a campaign in learning amounts to stopping the experiment before you have the result.
Exiting learning, on Meta and on Google
According to Meta's documentation, an ad set generally exits learning after about 50 optimization events over a 7-day period. If it does not get there, it moves to the “Learning limited” status: delivery continues, but the algorithm lacks the signals to stabilize.
On Google Ads, the mechanism is similar. According to Google's documentation, the “Learning” status appears in particular when a bid strategy is created, when campaigns are added to or removed from it, or when the conversions it optimizes for change. How long it lasts depends mainly on conversion volume and the conversion cycle, that is, the time between the click and the conversion.
In both cases, the practical conclusion is the same: low conversion volume lengthens learning, and learning that never ends signals a structure too fragmented for the budget.
Avoid restarting learning endlessly
On Meta, a significant edit restarts learning: a change of targeting, creative, optimization event, or bid strategy, adding an ad, a large budget change. Each edit of this kind resets the counters.
The costliest reflex is to pile adjustment upon adjustment on a campaign that is off to a bad start. It never has time to exit learning, its results stay unstable, which prompts yet more corrections. It is better to group changes together, then wait for learning to end before judging.
When a campaign stays stuck in learning, the levers are structural: consolidate ad sets that are too numerous, broaden the audience, increase the budget, or optimize for a more frequent event, higher up the purchase journey.
Frequently asked questions
How long does the learning phase last?
It depends on conversion volume. According to Meta's documentation, an ad set generally exits after about 50 optimization events over 7 days; on Google, the duration depends on conversion volume and the conversion cycle. A low-volume account therefore stays in learning longer.
What is the “Learning limited” status?
It is the status Meta shows when an ad set does not get enough optimization events to exit learning. Delivery continues, but performance remains unstable. The answer is structural: consolidate, broaden, or choose a more frequent event.
Should you change a campaign during learning?
As little as possible. Every significant edit restarts learning and prolongs the instability. If a change is essential, group all the corrections into one go rather than spreading them out.
Is a campaign in learning a bad campaign?
No: it simply cannot be judged yet. Its results reflect the algorithm's exploration, not the performance the campaign will reach later. Only stabilized periods can be usefully compared.