Meta Just Lowered the Advantage+ Learning Threshold to 25 Conversions, Here's Why That Should Make You Nervous
Meta cut the conversion threshold for Advantage+ Shopping campaigns from 50 to 25 per week, and rolled out Predictive Budget Allocation, which shifts spend in real time based on predicted conversion probability rather than distributing it evenly.
On paper, that’s good news: campaigns exit “Learning Limited” faster, and smaller advertisers who previously couldn’t hit the old volume requirement now qualify for the same automation larger accounts have used for years.
In practice, it means the algorithm is getting real budget authority on half the conversion data it needed before, and your Meta ads ROAS accuracy is the thing quietly absorbing that risk.
Here’s what actually changed, why “exited learning phase” doesn’t mean what it sounds like, and how to sanity-check the automation instead of just trusting the dashboard.
What the Threshold Drop and Predictive Budget Allocation Actually Do
The mechanics are straightforward. Advantage+ Shopping campaigns used to need 50 purchase conversions in a rolling 7-day window before Meta considered the campaign to have exited its learning phase and could optimize with confidence.
That threshold dropped to 25 in 2026, and Advantage+ App campaigns reportedly dropped even further, to 15.
- Meta’s Andromeda ad retrieval system, the engine now doing most of the ad selection work, is a large part of why the company felt comfortable making the cut: better signal extraction from smaller datasets, at least in Meta’s own telling.
Predictive Budget Allocation is the second half of the change.
Instead of spreading spend evenly across ad sets or waiting for a human to reallocate based on last week’s results, Meta now shifts budget in real time toward whichever ad set has the highest predicted conversion probability, with Meta reporting 8 to 15 percent ROAS improvement in early tests.
Combined, the two changes mean campaigns reach full budget authority faster and that authority moves money around more aggressively than it used to, all based on a data foundation that’s half the size it was a year ago.
Why “Exited Learning Phase” Doesn’t Mean “The Algorithm Has Enough Data”
Learning Limited status is a threshold, not a confidence score.
It tells you Meta’s system has decided it has enough data to stop labeling a campaign as unstable, not that the data is sufficient to make large, real-time reallocation decisions with low error.
- Twenty-five weekly conversions is a small sample by any standard statistical measure, and the rollout itself has been inconsistent enough that even close observers can’t fully agree on the current numbers.
One widely read changelog noted that some advertisers are seeing thresholds as low as 10 conversions over 3 days, while others still see the older requirements, with no consistent public confirmation of which applies where.
That opacity matters.
If you don’t know exactly what threshold your account is operating under, you can’t know how much signal is actually behind a given budget shift.
A campaign that “graduated” from learning on 25 conversions and immediately started reallocating aggressively under Predictive Budget Allocation is making confident-looking decisions off a data volume that would make most statisticians uncomfortable.
How to Run a Manual Holdout to Sanity-Check Advantage+ Decisions
The fix isn’t to turn Advantage+ off. It’s to verify what it’s doing against a number the platform doesn’t control.
- Hold a small percentage of budget or audience out of the automated allocation, even 10 to 15 percent, and track its performance against the automated segment using your own first-party, click-verified conversion data rather than Meta’s in-platform reporting.
- Compare campaign-level reported ROAS against blended MER over the same rolling window. If Advantage+ ROAS is climbing sharply while your overall marketing efficiency ratio stays flat, the gap is a signal the algorithm’s confidence is outpacing its actual accuracy. Our breakdown of platform-reported ROAS vs actual ROAS covers this comparison in more depth.
- Watch for reallocation frequency, not just direction. Predictive Budget Allocation reallocating budget several times a day off a 25-conversion base is a different risk profile than a weekly, human-reviewed shift, even if both land on the same eventual number.
- Run the holdout for at least two to three weeks before drawing conclusions, since a single week of divergence could just be normal variance at this sample size. Our guide to incrementality testing covers how to structure a holdout properly.
Watching for Correlated Overspend Across Multiple Accounts
If you manage more than one account, or an agency book of accounts, there’s a second-order risk worth watching for: correlated behavior.
When many accounts running the same Predictive Budget Allocation logic react to the same broad signal, whether that’s a seasonal demand shift, a competitor’s promotion, or a platform-wide trend Meta’s models pick up on, they can all shift budget in the same direction at roughly the same time.
That’s not a bug specific to any one account.
It’s what happens when a large number of independent decision systems are actually running the same underlying model on thinner data than before.
An agency managing several accounts in the same category should watch for spend moving in lockstep across clients that aren’t otherwise coordinated, since that pattern is a sign the automation is reacting to something structural rather than something specific to each account’s actual performance.
When to Intervene Manually vs. Trust the Automation
Neither extreme, always intervening or never intervening, holds up well under a lower confidence threshold.
A more durable rule of thumb:
- Trust the automation when your holdout test and blended MER are tracking consistently with reported ROAS over multiple weeks, and when reallocation patterns look proportional to real performance differences rather than erratic.
- Intervene manually when reported ROAS and blended MER diverge meaningfully, when reallocation frequency spikes without a clear driver, or when a campaign just exited learning on the lower end of the conversion range and hasn’t yet built a track record you can independently verify.
- Always keep independent, first-party attribution running underneath either decision. A lower confidence threshold makes independent verification more valuable, not less, precisely because the platform is now making bigger calls off thinner signal.
If you want to see how your Advantage+ campaigns are actually performing against verified, click-based conversion data instead of Meta’s in-platform reporting, book a live AdBeacon demo.
—-
FAQ
What is the new Advantage+ Shopping conversion threshold?
Meta lowered the requirement from 50 to 25 purchase conversions in a rolling 7-day window for Advantage+ Shopping campaigns to exit learning phase. Advantage+ App campaigns reportedly dropped further, to around 15.
What is Predictive Budget Allocation?
It’s a Meta feature that shifts budget in real time across ad sets toward whichever one has the highest predicted conversion probability, rather than distributing spend evenly or waiting for manual, periodic reallocation.
Does exiting Learning Limited mean my campaign has enough data?
Not necessarily. Exiting learning phase means Meta’s system has hit its own defined threshold, which recently got smaller. It doesn’t mean the underlying sample size is statistically robust for aggressive, real-time budget decisions.
How can I verify whether my Advantage+ campaign is actually performing well?
Run a manual holdout comparing a small held-back budget or audience segment against the automated portion, using your own first-party, click-verified conversion data, and compare reported ROAS against blended MER over several weeks rather than trusting either number in isolation.
Should I disable Advantage+ given the lower threshold?
Not necessarily. The automation itself isn’t the problem. The risk is trusting its output uncritically at a lower confidence level than before. Pairing it with independent verification lets you keep the speed benefits while catching decisions made on thin data.