You will be able to recognise the learning phase and avoid restarting it by accident.
Farah launched her Meta campaign on a Monday. On Tuesday each sale cost S$70. On Wednesday it cost S$35. On Thursday she made no sales at all and spent S$40. The figures are examples, but anyone who has launched a campaign will recognise the shape. Her instinct was to do something: lower the budget, change the audience, swap an image. She wrote the changes down, then made herself wait.
That wait was the right call. What she was watching was not a broken campaign. It was a new one, still working out who to show the ads to.
When a campaign or ad set starts, the platform knows very little about which people will respond to these particular ads. It has general knowledge about its users, and your audience settings and conversion history to start from, but it has to find out the rest by trying.
So in the first days it casts around. It shows the ads to different kinds of people, in different placements, at different times, and watches who converts. Each conversion teaches it something. Gradually it shifts delivery towards the people and places that work.
Meta and TikTok call this period the learning phase, and Ads Manager shows a learning status on ad sets going through it. Google shows a similar status when a smart bidding strategy is adjusting after a change. The names and the details differ by platform, but the idea is the same.
During learning, results are less stable and often more expensive than they will be later. The platform is spending part of your budget to find out what does not work. One day it tries a group that buys, the next it tries a group that does not, and your daily numbers swing.
That is why judging a campaign in its first few days is unreliable. Farah's Tuesday, Wednesday and Thursday were all the same campaign behaving normally. Averaged over the first week or two, the swings mean far less than any one of them did on the day.
If you react to every swing, you make it worse, for the reason in the next section. So decide in advance how long you will leave a new campaign before you judge it, and write that date down. Lesson 8.4, Write your budget and bidding plan, makes this part of the plan.
Some changes are large enough that the platform treats the ad set as new and starts learning again. Meta's help pages call them significant edits, and TikTok and Google have their own versions of the same idea.
Typical examples are a large change to the budget, a change to the audience or targeting, adding new creative or changing the existing ads, changing the bid strategy or a cost target, and changing the conversion event the ad set optimises for. Pausing an ad set for a while can also send it back into learning when it restarts.
Each of these resets some of what the system had learned, and the wobble begins again. That is why editing a campaign every day, the fourth mistake in lesson 6.4, Mistakes that waste a first budget, is so costly. A campaign that is edited daily never leaves the learning phase, so you only ever see its most unstable and expensive behaviour.
Small changes, such as fixing a typo in a headline, usually do not count. But the exact rules, including how big a budget change has to be before it counts, are set by each platform and change over time.
Each platform publishes a help page on how its learning phase works. It explains what the status means, roughly how much data the system needs before learning ends, which edits reset it and what to do if an ad set stays stuck in learning. Those details matter, and they change, so read the current page rather than relying on a figure from a blog post or an old course.
When you read it, look for three things. What has to happen for an ad set to leave the learning phase? Which edits does the platform say will restart it? And what does it suggest if an ad set cannot get enough conversions to finish learning, which often happens on small budgets or with rare conversion events?
The answer to the last question often points back to choices you made earlier in this course. Fewer ad sets so each gets more data, as lesson 2.3, Keep the structure simple enough to read, argued. A broader audience, as lesson 4.2, Broad targeting and why platforms now push it, suggested. Or a conversion event that happens more often, from lesson 2.1, The objective tells the platform what to find.
Farah's campaign settled in its second week, with each sale costing an example S$45 on average. She was glad she had not changed anything on Thursday.
In the activity below you will find and read the learning phase page in Meta's or TikTok's help centre, and write down what counts as a significant edit.
Read the learning phase page in Meta's or TikTok's help centre and write down what counts as a significant edit.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).