Why is my Meta ads performance so inconsistent?

If your Meta performance looks like a rollercoaster, one great week then two bad ones, and you can never work out which ads are your real winners, there’s usually one reason. It’s not your creative, and it’s not the algorithm having a bad month. It’s that the way you’re testing and managing ads is working against the algorithm. We see this pattern all the time across the ecommerce accounts we manage, and it happens in two ways.

Why is my Meta ads performance so inconsistent?

Both causes come down to the same thing:

1. Fragmented testing starves the algorithm of data. Ads tested one per ad set never get enough volume to give a reliable read.
2. Constant intervention resets the data it does have. Every time ads get turned off or swapped, the ad set goes back into learning.

Fix both and the account usually settles down within a couple of weeks: consistent delivery, CPAs you can trust, and winners you can actually identify and scale.

What’s wrong with testing one ad per ad set?

Here’s a setup we see a lot. A brand wants to test plenty of creative, so it launches each new ad in its own ad set, forces some spend into each one, and checks the in-platform CPA after a few days to decide what works.

On paper it feels scientific: every ad gets its own little test. And you do still get data this way. The problem is how much you can trust it.

For the algorithm to properly optimise an ad set, that ad set needs to exit the learning phase, and that takes around 50 conversions a week. Do the maths. If your target CPA is $100, one ad set needs about $5,000 a week, which is about $700 per day, just to exit learning.

Now, if you’re launching 20 ads a week, each in its own ad set, there’s pretty much no realistic budget that gets all of them through learning. So most of those ad sets stay in learning the whole time, and delivery in learning is a lot less stable. Some days an ad does well, some days it doesn’t, and over a short window you get a lot of random anomalies.

An ad might look like a winner off the back of a few lucky days. Another might look like a loser when it just didn’t get a fair run.

Why don’t my Meta ad winners hold up when I scale them?

This is a big part of where the inconsistency comes from. You take those winners, put them together and scale them up, and they may or may not hold up, because the read you got on them was taken while everything was still in learning.

It’s not that the data was fake. It just wasn’t stable enough yet to make those calls with confidence.

How should you structure Meta ad testing instead?

Consolidate it. Run one testing ad set, and fund it properly.

The formula is simple: your target CPA × 7. That’s the daily budget the ad set needs to get around 50 conversions a week.

Target CPA	Daily budget (CPA × 7)
$30	$210 per day
$50	$350 per day
$100	$700 per day

Then, instead of one ad per ad set:

1. Put a batch of 5 to 10 ads into the one ad set. Every ad in the batch should be a genuinely different concept: a testimonial, a before and after, a demo, a UGC piece.
2. Let them compete. The ad set is getting enough volume to read properly, so the algorithm pushes spend to what’s working, and because it can exit learning, the read is far more reliable.
3. Turn the losers off after about a week.
4. Bring in the next batch of fresh concepts to compete with the surviving winners.

It’s the same total budget, with much more stable data to make your calls on.

If you want to see where this testing ad set sits in the wider account, we cover that in the two-campaign Meta ads account structure.

Should you turn off a Meta ad with a high CPA?

This is the second way people fight the algorithm, and it’s even more common.

You look in the account and see your top-spending ad has a higher CPA than the ads next to it. Every instinct says turn it off and let the budget flow to the cheaper ones. But the algorithm optimises at the ad set level, not the individual ad level. When Meta ran tests on turning off high-CPA top spenders, the overall campaign CPA went up in basically every single test.

The ad you think is wasting money is usually doing assist work. It reaches people earlier in the journey who then convert through another ad, and the platform can only attribute the purchase to one of them. This is called the breakdown effect. That expensive-looking ad might be the exact reason your cheap ads look cheap.

Does turning ads off reset the learning phase?

Yes. Every time you turn ads off, or swap creative in and out, you reset the learning on that ad set. The account never gets a chance to stabilise.

That’s the other half of the rollercoaster. It’s not the algorithm being moody. It’s the account getting kicked back into learning every few days by constant changes.

When should you actually turn off a Meta ad?

Judge performance at the ad set level, and follow three rules:

1. If the ad set is hitting your CPA target, leave the ads alone, even if one of them looks expensive on its own.
2. Only turn off an individual ad when the ad set is trending above target and you can clearly see that one ad is dragging it: real spend, no conversions, over a week or more.
3. Clean out ads getting basically zero spend every week or two. That’s just housekeeping.

Otherwise, hands off. The less you interfere, the more stable the whole account gets, and it usually performs better too.

What else decides whether this works?

This fixes the stability side, but it isn’t the full picture. You still need the right overall account structure to plug it into, a creative pipeline feeding you fresh concepts every week, and attribution that tells you which winners are real beyond the platform numbers.

For the structure, start with the two-campaign Meta ads account structure or the ABO + CBO hybrid framework. If you’re working with a smaller budget, scaling Meta ads on a small budget shows how the same ideas apply from $100 per day.