From $90 to $27 CPA: A Real Week-by-Week Creative Testing Diary

"I just need to know it's actually working before I spend another dollar."

A founder said that to me on a call once, about six weeks into a build that was going fine but didn't look fine yet. And I get it completely. The honest answer was: it is working, you just can't see it in the numbers for another two or three weeks, and in the meantime it might look worse before it looks better.

That conversation is why I wanted to write this one as a diary. Most case studies skip the messy middle and jump straight to the happy graph. I'm going to walk you through an account turnaround week by week, the way we'd report it to a client, including the dip, the budget we were prepared to lose, and the exact moment consistent testing started to compound.

The numbers here are invented and anonymised to illustrate the rhythm, not a guaranteed result. But the shape is real. This is how these turnarounds actually go.

Week zero: the setup and the budget we were willing to lose

The account came in sitting around a $90 cost per purchase on a product that could comfortably support roughly $30 if we got it right. So the gap was real, and the maths didn't work at $90. That's the starting line.

First thing we did was strip it back. The previous setup had a dozen-plus ad sets layered with interests, lookalikes, a bit of everything, all splitting the budget into pools too small for Meta to learn from. We rebuilt to one clean manual sales campaign, broad targeting, and a single job: find the creative that works.

Here's the part most case studies leave out. Before we spent a cent, we agreed on a number we were willing to lose. We set the budget at a level the brand could spend for a full month and not feel it if it returned nothing. Call it a fixed daily amount, treated as the price of finding the answer, not a bet we needed to win this week.

I say this to every founder. The reason most people fail at Facebook ads is they treat it like a slot machine. They put a small amount in on Monday, it doesn't convert, they panic and pull it Tuesday. Meta never gets enough data to optimise, so it never works, so they pull it again. The willing-to-lose budget exists to break that loop. We're buying information for a month. Some of it costs money before it pays off.

Weeks one to two: the dip nobody puts in the case study

The first fortnight was slow, and I want to be honest about that because it's the bit that scares people.

We launched the first batch of creative. Week one returned a single purchase. Not a good week. If you were watching the dashboard hoping for instant proof, this is the moment your stomach drops.

We held. We didn't touch budgets, we didn't kill the campaign, we let it spend. Because a new account or a rebuilt one goes through a transition while Meta recalibrates around the new structure and the new ads. Early performance is noisy and usually understates where you'll land.

By the end of week two we had a handful more sales and, more importantly, we had data. Not winners yet. Just enough signal to read the creative. That's the trade in the early weeks: you're not buying sales, you're buying the information that tells you which way to push.

If a client only saw weeks one and two, they'd conclude it was a failure. The cost per purchase was still ugly, hovering up near where it started. The whole game is having the nerve to keep going when the number hasn't moved yet.

How we decide what to kill and what to keep

Before the week-by-week continues, here's the rule set we ran the whole way, because every kill and keep decision came from it.

We launched two to three new ad sets a week, each one a genuinely different idea, not a tweaked colour or a new font on the same hook. Big swings, every week, on a schedule.

Each week we made three moves:

• Kill the ones that barely spent and didn't convert. If Meta won't put money behind an ad, that's Meta telling you it can't find people who like it. Believe it.
• Kill the ones that spent well but came in above our control. Spend with a worse cost per purchase than your best ad set is spend in the wrong place.
• Never kill the control. Whatever ad set was carrying the spend at the lowest cost per purchase stayed on, untouched. You don't switch off the thing that's working to chase a maybe.

One read that saved us a few times: frequency. When an ad set spent very little but showed a high frequency, it meant the ad only appealed to a tiny pocket of people, so Meta kept showing the same few faces and couldn't scale it. Low spend plus high frequency equals a narrow ad. Knowing that stopped us mistaking a small retargeting-style winner for something that could actually grow.

And we judged everything on 7-day windows, never single days. Meta is consistently inconsistent day to day. A Tuesday with no sales next to a Monday full of them tells you nothing. The rolling week is the only number worth reacting to.

Weeks three to four: the first signs of life

Around week three the account started breathing.

A couple of the new ad sets began picking up spend on their own. Not a breakout, but the cost per purchase on the better ones was clearly under the account average, which had been stuck up near $90 all month. We killed the week's dead weight, kept the two showing promise, and launched the next three ideas.

Week four was where the discipline started paying. One of the fresh concepts took a meaningful share of spend mid-week and came in dramatically cheaper than everything before it. For a day or two we saw a cost per purchase less than half the account's month-long average. We left it alone and watched.

This is the moment the whole approach hinges on, and it's worth being precise about why it happened. We didn't get smarter in week four. We just kept putting genuinely new ideas in front of Meta every single week, and eventually one of them hit. Volume of real swings over time is what surfaces a winner. You can't predict which week it lands. You can only make sure you're still swinging when it does.

Weeks five to six: the winner takes over and the number drops

Once that concept proved it wasn't a one-day fluke, the account changed character fast.

Across the back half of the test, the winning ad set settled into carrying the majority of spend at a cost per purchase that pulled the whole account down with it. From a month sitting around $90, the rolling 7-day figure came down into the high $20s. Roughly a 70% cut in what we were paying to acquire a customer, off the back of one concept that didn't even exist in week one.

There were still ugly days in there. One Saturday spiked back up, the account briefly hit a spend limit and cut out mid-afternoon, normal Meta noise. But on the 7-day view the trend held. That's the payoff for judging on rolling windows instead of flinching at every red day.

Only once the cost per purchase was consistently where the product economics worked did we start scaling, and gently, nudging the budget up around 20% a day while performance held. Scaling is the reward for finding the winner, not a substitute for finding it.

What the diary actually teaches

Strip away the week numbers and three things did the work here.

A budget we were genuinely willing to lose, so we never made a panicked decision with money we needed back this week. The patience to sit through a transition dip that, in isolation, looked like failure. And a boring, relentless cadence of two to three real creative swings a week, judged on rolling windows, killing the dead and the over-priced and never the control.

None of that is clever. It's mostly just refusing to quit in weeks one and two, which is exactly where most accounts get switched off right before they'd have turned.

Where to from here

If your account is sitting at a cost per purchase that doesn't fit your economics, before you blow it up, ask one honest question: have you actually run consistent weekly creative tests on a budget you could afford to lose, for long enough to get through the dip? Most "it doesn't work" stories are really "we stopped in week two" stories.

This week-by-week loop is the exact thing we run for the brands we work with, and you can run a version of it yourself starting Monday. Pick a willing-to-lose number, commit to it for a month, launch two or three genuinely different ideas a week, and judge on 7-day windows. Then reply and tell me what week three looked like. That's usually where it gets interesting.