You had a winning ad set. You increased the budget to scale it. Performance fell off a cliff the next day. This is one of the most common and most avoidable mistakes in Meta advertising.
What the learning phase actually is
Meta's delivery system needs a minimum number of conversion events, generally around 50 per week per ad set, to exit "learning" and reach stable, optimized delivery. During learning, Meta is actively testing audiences, placements, and creative combinations to find its footing. Performance during this phase is inherently less efficient and less predictable than post-learning delivery.
Why a budget increase resets it
A significant edit to an ad set, and Meta's own guidance flags budget changes above roughly 20% as significant, can trigger re-entry into learning phase. The algorithm essentially restarts its exploration process for that ad set, which means a few days of worse performance while it re-stabilizes, right when you expected scaling to improve results.
The mistake most people make
Seeing a strong ROAS and immediately doubling the daily budget, then panicking two days later when ROAS drops and either killing the ad set or slashing the budget back down, which just triggers another learning reset. This start-stop pattern can keep an ad set permanently stuck in learning phase, never reaching the stable performance it's actually capable of.
What this looks like for any scaling brand
Whether it's a fashion brand pushing a bestseller or a skincare brand scaling a hero SKU, the pattern is identical: a working ad set gets a large, sudden budget jump, the account holder sees the dip and panics, and the campaign gets edited again before it ever had a chance to stabilize.
How to scale without the reset
Increase budget in smaller increments, generally recommended in the 10-20% range every few days rather than large jumps. Duplicate a winning ad set at a higher budget instead of editing the original, letting the original keep delivering on stable data while the duplicate builds its own learning phase in parallel. And give any change at least 3-4 days before making another one, most scaling mistakes come from reacting to day-two data instead of waiting for the algorithm to actually stabilize.