The short answer
The learning phase is the period after launch or a significant edit when Meta's delivery is still calibrating an ad set. Meta's guidance is still about 50 optimization events after the last significant edit. Creative volume interacts with it: adding ads counts as an edit, and on our board 54% of winners started within a week of another winner from the same brand, a sign that brands launch in batches.
What the learning phase is
Every time you launch an ad set or make a significant change to one, Meta's delivery system spends a period figuring out who to show it to. That is the learning phase. Costs swing more during it, and the results are a poor guide to how the ad set will perform once it settles.
Meta's help center gives the working number: an ad set should have about 50 optimization events since its last significant edit before its costs are stable. The event is whatever you optimize for. A purchase-optimized ad set needs purchases, not clicks.
Some advertisers have reported a lighter threshold of 10 events on certain purchase-optimized campaigns (Madgicx, 2024), without a matching change in Meta's general help pages. Meta's general documentation still describes the 50-event rule, so plan around 50 and treat anything lighter as a bonus.
If an ad set is unlikely to get there, Ads Manager labels it Learning Limited. That is a diagnosis, not a penalty.
What resets it
Meta's help center lists the edits that count as significant. The ones that matter most for creative work:
| Edit | Resets learning? | Notes |
|---|---|---|
| Adding a new ad to the ad set | Yes | The one most teams forget |
| Changing an ad's creative | Yes | Editing a live ad is an edit, not a new test |
| Changing targeting or the optimization event | Yes | |
| Pausing the ad set for 7 days or more | Yes | |
| Changing bid strategy | Yes | |
| Changing budget | Sometimes | Depends on the size of the change |
The first row is where creative volume and the learning phase collide. If your scaling ad set gets one new ad every two or three days, it may rarely be out of learning at all.
Winners arrive in batches
We can't see anyone's edit history, but we can see when winning ads started. In our October 9, 2026 snapshot (2,443 live Meta ads with $50k+ in estimated spend), we took every brand with two or more winners and checked how close each winner's start date was to its nearest sibling.
54%
of winners started within 7 days of another winner from the same brand
39%
within 3 days
21%
on the same day
4
median winners from one brand's busiest 30-day window (5+ winner brands)
That clustering is what batch launching looks like. Brands put several ads live together, then let them run. It is consistent with working around the learning phase: one significant edit for several new ads, instead of several edits in a row.
There is a second lesson in the same data. Ads that launched on the same day as a sibling reach Ad Radar's top tier (top 10% by winner score, live 21+ days) at 7.0%. Ads without a same-day sibling reach it at 11.0%. Same-day batches share the same ad set's learning, and Meta tends to pick one ad and feed it. Batches work, but expect one leader per batch, not three.
A batch in practice
Factor launched three ads on the same day, 117 days before our snapshot. Delivery picked one fast.
Anybody who says there's no shortcuts in keto hasn't tried Factor.
- Est. spend
- $4.1M
- Days live
- 117
- Format
- Video, 28s
- Variants
- 2
The batch leader: over $4M in estimated spend and in the top tier.
The other guys say that we're a bunch of faceless chefs, but that's because we're spending time in the kitchen.
- Est. spend
- $70K
- Days live
- 117
- Format
- Video, 30s
Same day, same offer, a different angle. Still live after 117 days, at under 2% of the leader's estimated spend.
The second ad is not a failure. It cleared $50k and is still running. But if Factor had judged the batch on week one, it would have seen one ad taking almost everything. That is how delivery behaves inside an ad set during and after learning, and why a separate testing lane matters for concepts you want a real read on.
Mistakes that keep ad sets in learning
An ad set stuck in learning is often not short of budget but of structure or patience. Each of these common patterns breaks one of the rules above:
| Mistake | Why it hurts | Fix |
|---|---|---|
| Adding one new ad every few days to the scaling ad set | Each addition is a significant edit | Batch new ads, or test them elsewhere |
| Editing the copy of a live winner | Restarts learning on the ad that was working | Launch the new copy as a separate ad |
| Ten ad sets on a budget that buys 40 purchases a week | Each ad set gets about four events | Consolidate to one or two ad sets |
| Pausing everything over a slow weekend | Pauses of 7+ days count as significant | Lower budget instead of pausing |
| Optimizing for purchases with very low volume | The event is too rare to exit learning | Optimize for an earlier event until volume grows |
| Judging a batch after two days | Costs are least stable early | Wait for the event count, then judge |
None of these is about the quality of the ads. Fixing them often recovers performance before any new creative is made.
How creative volume should work with learning
- Batch new ads. Launch four to six concepts together rather than dripping one in every few days. One significant edit, one learning period.
- Keep the scaling ad set quiet. Add new concepts in a testing ad set or campaign, then move winners over as new ads. Our guide to account structure shows the layout.
- Size the ad set to the 50 events. Divide weekly budget by target CPA. If the answer is well under 50, use fewer ad sets, a bigger budget per ad set, or an event higher in the funnel while you build volume.
- Expect one leader per batch. If you need every concept judged, give each a minimum spend in its own ad set. See creative testing budget.
- Don't edit live winners. A new hook is a new ad next to the old one, not an edit to it. Editing restarts learning on an ad that was working.
- Judge after learning, not during. Read results from the last significant edit forward, as Meta suggests, and wait for the event count before calling it.
Clean conversion data is the other half of this. The learning phase counts the events Meta receives, so missing or duplicated purchases distort it directly. That is covered in Conversions API and signal quality.
When you research competitors, start dates are visible for every ad. Brands that put several ads live on the same day are showing you their batch size, and on Ad Radar you can follow a brand and filter its winners by days live to spot each batch.
Figures marked as estimated spend come from Ad Radar's model of engagement on public Meta Ad Library ads. They are estimates, labeled as such, and are best used to rank ads against each other.