What is the learning phase on Meta?

Meta's delivery system learns from every impression. Each time it shows an ad, its predictions of relevance get more accurate, and the more it shows an ad, the better it gets at choosing where, when and to whom to show it. Meta calls the period when the system still has a lot to learn the learning phase.

Learning starts when someone creates a new ad or ad set, or makes a significant edit to an existing one. The system then explores the best way to deliver the ad set for the optimisation event it was given. Meta says performance is less stable during this period, cost per result is usually higher, and results are not necessarily a guide to future performance.

The system never stops learning, but the learning phase ends. Ads Manager marks it with the word Learning in the Delivery column of the ad set.

Meta says ad sets exit the learning phase as soon as they can deliver stably, which usually happens after about 50 results in the week after the last significant edit. Its Learning limited page phrases the same threshold as about 50 optimisation events in the week after the last significant edit. The count belongs to each ad set and starts again after each significant edit.

A result here means the optimisation event the advertiser chose for the ad set. An ad set optimising for purchases needs about 50 purchases in a week. One optimising for add to basket needs about 50 add to basket events, which is why Meta suggests moving to a more frequent event when an ad set cannot get enough.

Shops ads have their own threshold. Meta says they need at least 17 purchases through the website and 5 through Meta for learning to complete. In Ads Manager, the Last significant edit column shows when the count last restarted, and the Results column shows progress since then.

Every edit affects delivery a little, but only a significant edit sends an ad set back into learning. Meta lists these changes:

Meta's budget example is a 1% increase, which is unlikely to restart learning, against a tenfold increase, which may. With Advantage+ campaign budget, changing the campaign budget or switching the campaign bid strategy can send several ad sets back into learning at once. An edit made at ad set level affects only that ad set. Meta's developer documentation adds that migrating an old Advantage+ shopping campaign into the new Advantage+ sales campaign structure forces it into learning.

Adding a new ad counts as a significant edit. Because the count runs from the last significant edit, adding new ads in batches restarts a live ad set less often than adding them one at a time.

If an ad set is unlikely to get about 50 optimisation events in the week after its last significant edit, Ads Manager shows Learning limited instead of Learning. Meta says this is not a penalty. It means the budget is not being spent effectively because the delivery system cannot optimise with the current setup. The status can appear any time after an ad set is created or significantly edited.

Meta lists these causes:

• An audience that is too small.

• A budget too low to get enough optimisation events.

• A bid or cost control set too low, which Ads Manager flags in a tooltip on the status.

• High auction overlap. When several ads from the same advertiser enter one auction, Meta enters only the one with the highest total value, so overlapping ad sets keep each other out of auctions.

• An optimisation event that happens rarely.

• Running too many ads at the same time.

Meta's fixes follow from the causes. Combine ad sets and campaigns, widen the audience, raise the budget or the bid control, or choose a more frequent optimisation event, such as add to basket instead of purchase. An ad set that then collects enough optimisation events since its last significant edit moves from Learning limited to Active.

An ad set in learning spends money while the delivery system is still exploring. Meta says ad sets are less stable and usually have a higher cost per action during this period. Once an ad set exits, delivery is stable enough that its results start to predict what comes next.

Meta's advice is about giving each ad set enough data. It recommends waiting to edit until an ad set has left learning, avoiding edits without a clear reason, avoiding high ad volumes and setting a realistic budget. Meta says a very small or inflated budget gives the system an inaccurate signal about whom to optimise for. Combining similar ad sets also combines what the system learns about them.

Structure follows from this. For example, ten ad sets that each get 10 purchases a week all stay in learning, while one ad set with the same total of 100 purchases a week can exit. Meta says the same about ads, since running too many at once means each delivers less often and fewer exit learning. The testing budget guide turns this into budget numbers for tests.

Meta also says advertisers should not try to avoid the learning phase completely, because testing new creative and strategies is how performance improves over time. The aim is to restart learning only when an edit has a reason behind it.