Meta learning phase: what really happens to a new ad set in its first week, from real online stores

What the learning phase is, and what we measured

The learning phase is the period in which Meta does not yet know who to show a new ad set's ads to, so its results are unstable. According to Meta, an ad set leaves the learning phase after about 50 conversion events in a week, which for an online store means 50 purchases. An ad set that does not get there is labelled "learning limited", and the usual advice is to consolidate ad sets, raise the budget, and touch nothing for a week.

I run the team that launches campaigns for online stores, and these decisions come back every week: give a new ad set more time or close it, raise the budget now or wait, and what to do with an ad set that has been stuck in "learning limited" for a month. We wanted to know what our own data says about these rules, not what Meta says about them.

We took every Meta ad set in the Israeli online stores we work with, from April to October 2026, and followed each one from the first day it spent money. The sample: 12 stores, 185 ad sets that started running in that period and spent at least 1,000 shekels over their life, 764 thousand shekels of cumulative spend and 9,673 purchases. Purchases are what Meta reports on the ad set, not the store's revenue, so every figure here compares an ad set with itself, and none of it is ROAS. A week in an ad set's life counts only when it is full: at least six active days, 100 shekels and 1,000 impressions. Each metric is computed per ad set against its own first week, then the median is taken within the store, and then the median across the stores.

50 purchases a week: almost no ad set gets there

We started with the rule itself. Out of 816 full active weeks of ad sets, only 1.3% reached 50 purchases. One week in 74. The median ad set brought 5 purchases a week, and in 71% of the weeks it brought fewer than ten.

At the ad set level it looks like this: of 185 ad sets, 3 reached 50 purchases in their first week, and 92% never got there in any week of their life. Even the lower bar, 20 purchases a week, was ever crossed by only 58 ad sets.

"Learning limited" is not a fault in the account. In an Israeli online store with an Israeli store's budget, it is the normal state of most ad sets, most of the time.

That changes the question. If Meta's threshold is relevant to one week in 74, you cannot run campaigns by "waiting for learning to finish". You need to know what really happens to an ad set that lives below the threshold, and that is what we looked at next.

The first week was not the expensive one

The common assumption is that an ad set's first week is a week of burning money: Meta learns, and we pay for the lesson. In our data that did not happen. At the median, the cost per purchase in the second week was 115% of the first week, 111% in the third and 110% in the fourth. Weeks 2 to 4 together cost 14% more than the first week at the median, and 6% more when weighted by money.

Our explanation is the same one we saw at the level of the single ad, in the article on ad fatigue: a new ad set first meets the people who were waiting for the product, and then everyone else. Meta is not "learning" at our expense in the first week. It sells to whoever is easiest to sell to, and then the work gets harder.

The management conclusion is sharp: a first week with a high cost per purchase is not "still learning". It is usually the best week that ad set will ever see. When a new ad set does not work in its first week, we do not give it another week to "leave the learning phase". We change something, the audience, the offer or the ads, or we close it.

What Meta's threshold does measure: stability

If the first week is not expensive, what is different about an ad set that brings 50 purchases a week? In the data, the answer is stability. We measured how much the same ad set's cost per purchase moves from one week to the next, by the number of purchases it brings.

An ad set with fewer than five purchases a week moved 46% from week to week at the median. An ad set with 20 purchases or more moved 26%. 1.8 times less. That does not mean the bigger ad set is cheaper, only that its number can be trusted. When an ad set with four purchases a week shows a good cost per purchase, half of that is luck, and it can be gone next week.

So for us, the decision on an ad set is not measured in days but in purchases. An ad set that has collected 20 purchases is an ad set we can judge, even if it took two weeks. An ad set with three purchases after a week has not said anything yet, for better or worse.

And what about consolidating ad sets? We checked that too: in the weeks when a store ran few ad sets at once, its cost per purchase was 98% of its own median, and in the weeks with many ad sets 103%. Almost the same. At the median, a store ran 5 active ad sets a week, and the largest one took 36% of the weekly budget. Consolidation buys stability and a number you can read, not a lower cost per purchase. That is a good reason to consolidate, but it is the right reason.

What happens when you raise the budget

The second rule everyone knows: never change a budget by more than 20%, because it sends the ad set back into the learning phase. We took 478 pairs of consecutive weeks of the same ad set, and grouped them by what happened to the spend between the first week and the second.

The table says three things. First: even an ad set nobody touched got more expensive the following week, by 11%. That is the ageing we saw in the previous chart, and everything else has to be read against it. Second: raising the budget worked. Across all the ad sets whose spend rose by more than 20%, 77% more money brought 49% more purchases, at a cost per purchase 19% higher, that is 8 points above an ad set nobody touched. The increase did not "reset" the ad set. It bought more purchases, a little more expensively.

The third is surprising: the moderate increase, 20% to 50%, was the worst. 34% more money brought only 11% more purchases. Doubling, on the other hand, brought 74% more purchases for 103% more money. Our explanation, from working on the campaigns: a big increase happens when there is a reason, a sale, new stock, an ad set that has proven itself. A small increase is often a 20% nudge every week on a middling ad set, with no new reason. The budget goes up, the offer stays the same.

What we do with this

This is how we manage ad sets after seeing these numbers.

1. Judge by purchases, not by days. A new ad set is reviewed when it has 20 purchases, not when a week has passed. Below that the number moves 46% from week to week, and there is nothing to decide on.
2. A bad first week is not "still learning". The cost per purchase in weeks 2 to 4 was 14% higher than in the first week. An ad set that did not work in its first week will not improve on its own. Change the audience, the offer or the ads, or close it.
3. Ignore the "learning limited" label. 92% of ad sets never reached 50 purchases a week. If every campaign in the store carries the label, the label says nothing about the store.
4. Consolidate for a readable number, not for price. Weeks with few ad sets and weeks with many cost the same. But an ad set with 20 purchases a week gives a number you can trust, and one with four does not. Fewer ad sets, each with enough purchases to know what is happening in it.
5. Raise when there is a reason, then raise properly. Doubling brought 74% more purchases. A 34% raise brought 11%. A sale, new stock or an ad set that has proven itself is a reason. "The week was fine" is not.
6. Plan for the cost to rise, even untouched. An ad set nobody touched got 11% more expensive the following week, and an ad set in its sixth week cost 120% of its first. A new ad set, with a new audience or offer, has to be ready before the old one gets expensive, not after. What gets more expensive in the market itself, rather than in the ad set, is in the article on how much Facebook ads cost.