There's a pattern we see in almost every ad account we inherit. The campaigns are set up fine. The creative is fine. But performance is stuck, and the reason isn't any single setting. It's that nobody let the algorithm finish learning before they started tinkering.
If you understand the learning phase, you understand why most ad accounts leave money on the table. So let's walk through it.
What the learning phase actually is
Every time you create a new ad set, or make a significant change to an existing one, Meta enters a calibration period called the learning phase. During it, the delivery system is actively experimenting: testing which audiences, placements, times, and creative combinations produce the result you asked for. It's expensive and volatile, because the algorithm is spending your budget to gather information.
To exit the learning phase and stabilize, an ad set needs roughly 50 optimization events, purchases, leads, whatever you're optimizing for, within a 7-day window. Until it hits that, delivery stays unpredictable and your cost per result bounces around.
The mistake that quietly kills accounts
Here's the trap. Performance looks shaky in the first few days, because it's supposed to, so the account manager panics and starts changing things. New audience. Different creative. Budget bumped up. Optimization event switched.
Every one of those is a significant edit, and every significant edit resets the learning phase. The algorithm throws away everything it learned and starts over, spending your budget to re-discover signals it had already found. The account never stabilizes because it's never allowed to.
Meta's Andromeda update, fully rolled out by early 2026, made this worse in one specific way: it tightened what counts as a "significant" edit and extended typical learning phase duration from 4 to 7 days out to 7 to 14 days for many accounts. Changes that used to be safe now trigger a reset. Patience matters more than it ever has.
Why "just leave it alone" is a strategy, not laziness
The single most valuable thing we do in a lot of accounts is stop touching them the wrong way. That doesn't mean set-and-forget. It means:
- We feed the algorithm clean, qualified conversion events through the conversion API, so it optimizes toward real revenue instead of cheap clicks.
- We give ad sets enough budget to actually hit 50 events in a week. If your target cost per result is $30, that's roughly $214 a day per ad set. Underfund it and it never exits learning, no matter how long you wait.
- We batch our changes and make them deliberately, at the campaign level, instead of nudging live ad sets every day and resetting the clock.
Structure decides how fast it learns
The other half of the equation is campaign structure. Under Andromeda, Meta uses creative signals and engagement behavior, hook rates, completion rates, to find your buyer across a broad pool, rather than relying only on the audience you defined manually.
That means the structure you launch with, how campaigns are consolidated, which objectives you choose, where the budget sits, decides how quickly the algorithm finds the right people. Most accounts we inherit are structured in a way that fights the algorithm instead of feeding it. Fixing the structure often does more than any single creative ever could.
The takeaway
If your ads underperform, the problem usually isn't that you need a new agency with a secret creative formula. It's that the account has never been allowed to stabilize, and it's structured to make learning slow.
Respect the learning phase. Fund it properly. Feed it clean conversion data. Structure it to help the algorithm instead of fighting it. That's less exciting than "we'll 10x your ROAS with one weird hook," but it's what actually produces a return that holds.