Two campaigns: testing and scaling
Scaling holds the ads that have already earned their place. Testing is where new ideas go, and we expect most of them to fail.
Scaling returned about 0.75 in settled cash on A$3,896 of spend and 93 buyers. It isn't paying for itself yet, and we're still tuning it.
Testing returned about 0.41 on the same estimate. We don't play it safe there: we spend some money looking for an offer far better than the one we have.
Of 18 test ad sets, 10 never made a sale. One that did was an angle I wouldn't have picked: keep your job and go build software. Buyers at A$27 each, against A$65 for testing overall. That's 5 buyers, too few to call it.
Estimate. Meta reported 0.65 for scaling. Across the whole funnel our bank received 16% more than Meta counted. Stripe doesn't record which campaign a sale came from.
What Meta's algorithm is doing with your ad
Nothing magic. It runs on every ad in every auction, so it has to be cheap. It starts from your account's history, basic tags, a rough look at the image. Then it watches early reactions and backs whatever got the first good one.
Hand out ten posters in a city, and if the first one gets a nod, the effort moves to the first one. If those first people weren't your best buyers, you get more people like them, and a better group never sees the ad.
Testing changes who sees an ad first. Most of those groups will be worse. You pay for those to find the one that's better.
Ads sell as a group
I put about five ads in an ad set. One says we build software, one asks if you knew you could build it yourself, the last is the specific offer. Meta credits nearly every sale to the offer ad. When that ad ran on its own, its sales dropped to zero.
So the Sheet grades the ad set, not the single ad. The ads that don't get the credit are still doing part of the work.
A lucky start is not a winner
One ad took three sales from about its first six views, on about a dollar of spend. Roughly $300 came back. I moved money onto it, spent another $500, and didn't get another sale.
Run hundreds of ads and that will happen a few times by chance. The fix is a rule you set before launch, so a lucky start can't talk you into anything.
The kill rule
2× before you judge.
Give every ad set at least 2× your target cost per buyer before you kill it. An ad twice as good as target dies early only 2% of the time at 2×, and 13% of the time at 1×.
On the Labs book, target A$35, that's A$70 per ad set, spread over about five days at around A$14 a day.
Modelled, not measured. It assumes sales arrive at a steady rate per dollar spent.