The Meta ads learning phase is the single most misunderstood part of running a mobile app account, and the misunderstanding is expensive. Teams see the word "learning" on an ad set, assume something is broken, and start editing. Every edit sends the ad set back to the start. Weeks later they are still in learning, still blaming the algorithm, and still resetting the very process they are waiting on.
The learning phase is not a bug and it is not a punishment. It is the period where Meta's delivery system works out who to show your ads to and how to buy for you efficiently. Under the Andromeda engine that now runs delivery, that calibration matters more than it used to, not less. Get it right and your costs settle. Keep interrupting it and they never do.
This is the operational guide: what the Meta ads learning phase actually does, why iOS apps need to be more patient than the defaults suggest, the five edits that quietly reset it, and how to fix learning limited when an ad set is stuck.
Key takeaways
- The learning phase is the calibration period where Meta stabilises delivery, and it exits at roughly 50 optimisation events in a seven-day window.
- On iOS, plan for seven to ten days, not three to four, because SKAdNetwork postback delay means the seven-day clock counts events it can see, not events that happened.
- Five edits reset learning: targeting, creative, optimisation event, bid strategy, and a large budget change. Treat every edit as a reset until proven otherwise.
- Learning limited is a structural problem, not a waiting problem. The fix is consolidation, not patience.
- Under Andromeda, fragmenting your signal across too many narrow ad sets is one of the fastest ways to stay stuck in learning.
What the Meta ads learning phase actually is
The Meta ads learning phase is the period, at the start of a new ad set or after a significant edit, during which the delivery system is still working out how to get you results efficiently. Performance is more volatile and cost per result is usually higher, because the system is exploring rather than exploiting. Meta considers an ad set to have exited learning once it has recorded around 50 optimisation events within a seven-day window. Below that, it does not have enough signal to settle.
The important word is "optimisation events". The clock does not count impressions or clicks. It counts the specific event you told the ad set to optimise for, whether that is an install, a trial, or a purchase. The rarer that event, the longer the phase takes, which is the root of most learning problems on app accounts.
How Andromeda changed the equation
Andromeda is Meta's retrieval engine, rolled out globally through 2025 and representing a roughly 10,000 times increase in model complexity at the retrieval stage. It narrows billions of eligible ads down to a small set of auction candidates in milliseconds, weighing predicted action rate, bid, and the user experience of the ad. All of that still depends on a clean, stable conversion signal, which is exactly what the learning phase produces.
The practical shift is that the old habits now cost more. Splitting your budget across a dozen narrow ad sets, or editing constantly to chase a bad day, fragments the very signal Andromeda is trying to read. If you want the deeper mechanics, we covered them in our guide to Meta's Andromeda algorithm. The short version: give the system fewer, stronger signals and it learns faster.
Why iOS apps need 7 to 10 days, not 3 to 4
On a web account with an instant conversion event, an ad set can clear learning in a few days. Mobile apps on iOS cannot rely on that timeline, and the reason is SKAdNetwork. Apple's attribution framework delivers install postbacks on a delay of up to 72 hours, and often in coarse, batched form. The seven-day learning window can only count the events it has actually received, so a real install today might not register for two or three days.
That lag stretches the calendar. An iOS app set optimising for installs realistically needs seven to ten days to exit learning cleanly, and any judgement you make before then is being made on incomplete data. The same lag is why single-day performance reads are so misleading, a theme that runs through everything from diagnosing creative fatigue to judging whether a fresh concept is working. Patience is not a virtue here, it is a measurement requirement.
The five edits that reset learning
Every significant edit puts the ad set back into learning and throws away the calibration you already paid for. These are the five to respect:
- Targeting. Changing the audience, adding or removing detailed targeting, or swapping the lookalike source all reset the phase.
- Creative. Adding a new ad or changing the creative in a meaningful way counts as a reset, though editing ad text without touching the asset usually does not.
- Optimisation event. Moving from installs to a deeper event, or vice versa, restarts learning because you have changed what the system is aiming at.
- Bid strategy. Switching between lowest cost, cost cap, and bid cap, or changing the cap itself, resets delivery.
- Large budget changes. A big jump or cut in budget can reset learning. Smaller, staged changes of around 20 percent are the safer way to scale without paying the reset tax. Pausing an ad set for seven days or more also sends it back into learning.
Learning limited, and how to fix it
Learning limited is different from being in learning. It means the ad set will never reach 50 events a week at its current setup, so delivery stays unstable indefinitely. Waiting does not fix it, because the maths does not work. The fix is almost always to consolidate and simplify.
Merge overlapping ad sets so the conversions pool into one place rather than splitting across five. Raise the budget if the conversion event is starved. If your deeper event fires too rarely to sustain learning, move the optimisation up the funnel to installs until volume supports something deeper. Under Andromeda this consolidation is not just a learning fix, it is how you feed the system the concentrated signal it rewards. Fewer, better-fed ad sets beat many starved ones every time.
Frequently asked questions
How long does the Meta ads learning phase last?
Meta exits the learning phase once an ad set records roughly 50 optimisation events within a seven-day window, so the calendar time varies with your budget and how frequent your conversion event is. For a mobile app optimising for installs on iOS, plan for seven to ten days rather than the three to four you might expect on web, because SKAdNetwork postbacks arrive on a delay of up to 72 hours and the seven-day clock only counts events it can actually see.
What does learning limited mean on Meta ads?
Learning limited means the ad set is not generating enough optimisation events to ever hit the 50-per-week threshold, so Meta cannot stabilise delivery. It is a structural signal, not a temporary one. The usual causes are a budget that is too low for the conversion event, too many ad sets splitting the same conversions, or an optimisation event that happens too rarely, and the fix is almost always consolidation rather than patience.
What edits reset the Meta ads learning phase?
Any significant edit resets learning: changing the audience or targeting, editing the creative or adding a new ad, changing the optimisation event, changing the bid strategy or cost controls, and making a large budget change. Pausing an ad set for seven days or longer also sends it back into learning. Minor changes, like editing ad text without swapping the creative, generally do not.
Should I optimise for installs or for a deeper event during learning?
Optimise for the event you can actually feed. A deeper event like a trial or purchase is a better business signal, but if it fires too rarely you will sit in learning limited forever. Many apps start on installs to exit learning cleanly, then move to a deeper event once volume supports 50 of those events a week. The right answer depends on your spend, not on which event sounds better.
Does the learning phase still matter under Andromeda?
Yes, and arguably more. Andromeda is Meta's retrieval and delivery engine, and it still needs a stable conversion signal to learn who to show your ads to. What changed is that constant edits and narrow ad sets now hurt more than they used to, because they fragment the signal Andromeda depends on. Respecting the learning phase is part of giving the system the clean signal it wants.
Want this run for you?
Most learning phase problems are self-inflicted: too many ad sets, too many edits, and too little patience with the SKAdNetwork clock. The accounts that get it right build a stable structure, feed it a clean signal, and then leave it alone long enough to settle. That discipline is part of the work we do on our performance creative agency engagements.
If you run a mobile app spending real money on Meta and your ad sets never seem to leave learning, apply to work with us. We take a small number of mobile app clients per quarter.