What resets the Meta Ads learning phase: changing the budget and adding creatives

Translated from the Japanese original on mureo.jp.

When you run Meta ads on a small budget, you run into a situation where Ads Manager shows “Learning” or “Learning limited” and it never clears. You raise the budget, thinking more money will speed up learning, and since the response is still weak, you add creatives too. Both of these moves are, in fact, changes that send learning back to the start.

We did exactly this, in this order, with our own Meta ads. We doubled the daily budget from ¥1,500 to ¥3,000, and the next day we added two creatives. We kept the numbers from that time, so here we write down what happened, how much of it can be put down to learning starting over, and where that reading stops.

One caveat up front: at the time of writing, we had not been able to confirm in the original Help Center text the detailed conditions under which Meta treats a change as resetting learning. We treat those conditions as what is commonly known, and keep them separate from the facts we did confirm.

What the learning phase is, and when it ends

The learning phase is the period during which Meta’s delivery system is working out how to optimize an ad set. During this period delivery is unstable and costs move a lot. The widely cited rule of thumb is that learning ends once about 50 of the events you optimize for occur within 7 days.

The point to note here is that only “the event you optimize for” is counted. If you optimize for link clicks, 50 clicks is enough. If you optimize for inquiries (leads), though, 50 means 50 inquiries in a week. For a business that gets only a few inquiries a month, this condition cannot be met from the start.

The “Learning limited” status appears when delivery is not expected to reach this condition. It does not mean the ad is broken; it means there are not enough signals for learning to use.

What makes learning start over

The changes most widely cited as resetting learning are the following. None of these were confirmed in Meta’s Help Center; they are what various explainers have in common.

Changing the optimization event: for example, switching from leads to link clicks
Changing targeting: adding or removing audiences
Adding or replacing creatives: adding a new ad
A large budget change: a change of more than 20% at once is the figure often given, but we could not confirm an official threshold
Pausing for 7 days or more: learning starts over when you resume

What matters is that raising the budget and adding creatives both look like obvious improvements. When the response is weak, these are the two things people touch first, and both send learning back to the start.

What happened in our account

Our Meta ads were optimizing for pixel leads, delivering with an ad set daily budget of ¥1,500. We made changes in this order.

Changed the daily budget from ¥1,500 to ¥3,000 (August 11)
The next day, August 12, added two creatives
Over this period, Meta recorded 1 lead in 8 days

Because the two changes came back to back, learning started over once, and then started over again before it had finished. On top of that, the signal, leads, came to only 1 in 8 days, two orders of magnitude short of the 50-a-week guideline. This delivery was never in a position to finish learning either way.

CPM (cost per 1,000 impressions) moved like this over the period.

August 4 to 6: ¥4,212
7-day average, August 4 to 10: ¥6,208
August 10 alone (daily budget still ¥1,500): ¥8,037
August 11 to 12 (daily budget ¥3,000): ¥10,346

At first we read this as “doubling the budget doubled the CPM,” because we compared the 7-day average with the most recent figure. Laid out day by day, CPM had already risen to ¥8,037 the day before the budget change, and the increase added by doubling was about 28%, from ¥8,037 to ¥10,346.

This correction is the caveat for reading the learning phase. If you put a CPM increase down to the budget change alone, you get the cause wrong. In this case, CPM had been rising since before the budget change, and we read it as more likely that lead optimization itself, with so few signals, kept buying a narrow audience at high prices. This is, however, an observation from one account, and we cannot say other accounts would behave the same way.

What to check before a change when running on a small budget

From this experience, we now check the following four things before making a change.

How many times a week the current optimization event occurs: if it falls far short of 50 a week, adding budget or creatives will not move learning forward. First consider whether to switch the optimization event to an earlier action that happens more often, such as a landing page view.

Change one item at a time: if you change the budget and creatives at the same time, or one right after the other, you cannot tell later which one caused a change in the numbers. Because we changed them back to back, we misread the CPM increase as caused by the budget.

Write down the date of each change: if there is no record of what was changed when, you cannot check afterward whether learning started over. Keep the change history separately rather than leaving it to the platform’s screens alone.

When raising the budget, also watch daily spend behavior: right after a budget increase, there are days when spend exceeds the setting (Why ads spend more than the daily budget). Several causes of noisy numbers overlap, so do not judge all at once.

If you have handed ad ops to an agent, the risky part is not running these changes but “how to read the numbers after a change.” Whether a change went in as intended can be checked with the steps for checking after a change, but whether learning started over is a separate question, and it does not appear in the agent’s report.

If you run a small budget and want to sort out, for your own account, where you are stuck in the learning phase and which changes to hold off on, get in touch with your current optimization event, the changes you made in the last two weeks, and the number of target events per week.

Where people tend to trip up in the learning phase

Reading noisy numbers as a bad ad: costs move a lot during learning. Do not judge which creative is better until learning has finished.

Adding budget because the response is weak: if the cause is a lack of signals, adding budget will not help. It only adds more resets.

Adding many creatives at once: each new ad adds to what has to be learned. On a small budget, it is easier to read the results if you keep the number down and swap them in one at a time.

Even just resuming after a pause has an effect: resuming an ad set that has been paused for a long time can send learning back to the start.

Frequently asked questions

How many days does the Meta Ads learning phase take to finish?

It is decided not by days but by the number of optimization events. The general guideline is about 50 in 7 days. If there are not enough events, it will not finish however many days you wait, and the status stays at “Learning limited.”

Does increasing the budget reset learning?

It is often explained that a large change at once resets learning, and a change of more than 20% is the figure commonly given. We could not confirm an official threshold, so the safe course is to change it gradually (for example, in small steps every few days). We doubled ours and went through a reset.

Does adding creatives affect learning?

Adding a new ad is generally treated as a change that resets learning. Testing in a separate ad set, rather than adding to an existing ad set that is delivering, does not disrupt the original ad set’s learning.

What should I do when it says “Learning limited”?

Check how many times a week the event you optimize for occurs. If it is too few, switching the optimization event to an earlier event that happens more often is more realistic than raising the budget. Changing the event itself resets learning, though, so write down the date of the change and leave everything else alone for a few days.

Is it a mistake to optimize for inquiries on a small budget?

It is not a mistake, but it is hard to reach 50 a week. You need to read the results on the assumption that you will not reach it and that learning will not stabilize. Our lead optimization had 1 signal in 8 days, which gave us nothing to judge costs on.

Summary

The Meta Ads learning phase lasts until a certain number of optimization events accumulate, and it starts over every time you make a large budget change or add creatives. We added creatives the day after doubling the budget, and with only 1 signal in 8 days, we reset learning twice in a row.

Reading the worsening numbers as the budget’s fault was premature. Laid out day by day, CPM had been rising since before the budget change. If you check the number of target events and limit yourself to changing one item at a time before making a change, you avoid misreading the cause later.