There's an assumption behind most ad accounts that rarely gets examined:
"If I optimize for purchases, Meta will optimize for my business."
That sounds reasonable. The platform is sophisticated. It uses machine learning. It talks about value optimization and conversion modeling. Surely it understands what's good for you.
It doesn't.
What the System Actually Computes
Meta's engineering papers describe exactly what happens when an ad slot opens. The system scores each candidate ad using a formula that combines predicted click probability, predicted conversion probability, and bid value. The GEM paper — Meta's core ad prediction model — describes a system built to estimate the likelihood that this user, seeing this ad, in this moment, will take the action the advertiser selected.
If you selected "Purchase," the system scores every available impression based on how likely that user is to purchase. The impression with the best combination of predicted probability and bid value wins the auction.
Not predicted profit. Not predicted margin. Not predicted lifetime value. Predicted probability of the event you chose.
This is not a flaw in the system.
It is what the system is designed to do. Meta's engineering papers describe an infrastructure built to evaluate billions of impressions per day against predicted event outcomes. That is a genuinely remarkable piece of engineering.
But it is solving Meta's problem — allocating ad impressions efficiently across millions of advertisers — not your problem, which is running a profitable business.
What Meta Cannot See
Meta does not know your business model. It doesn't know whether you are subscription or one-time purchase, high margin or razor thin, DTC, wholesale, or B2B, capital constrained or cash rich.
It doesn't know your cost of goods. It doesn't know your shipping or fulfillment costs. It doesn't know your overhead. It doesn't know your retention curve.
It knows something much narrower: for this impression, what is the predicted probability of the event you selected?
Meta optimizes for events. You operate a business.
Those are not the same problem.
An important nuance: value optimization gets closer — but not close enough.
If you're running value optimization campaigns and passing accurate revenue data through the pixel or CAPI, the system does orient toward higher-value conversions rather than pure purchase probability. This is real and worth using.
But even with value optimization, Meta is optimizing toward the revenue signal you send it. It still cannot see your margin on that revenue. A $200 order with 60% margin and a $200 order with 8% margin look identical to the system. A customer who will reorder five times and a customer who will return the product and never come back generate the same purchase event.
The system can optimize toward the value signal you give it. It cannot optimize toward the economics behind that signal. That's your job.
If Meta Doesn't Know Your Economics, You Must
If the platform can't see your margins, then you need to anchor performance to something it cannot see: your benchmarks.
At minimum, every advertiser should know their break-even CPA and their target CPA. If you generate leads, the same applies to CPL.
Break-even is not a guess. It is the acquisition cost at which you neither make nor lose money after COGS, shipping, payment processing fees, refunds, and variable costs. Target CPA reflects your margin goals on top of that.
Meta does not know these numbers. Left unconstrained, it will pursue conversion probability — not profitability. Your budget and bid strategy are the only tools you have to impose your economics on a system that cannot see them.
🖼 Asset needed — screenshot: Breezeway KPI Calculator. Caption: "Your break-even and target CPA should be calculated from your real economics — not guessed from yesterday's ROAS."
One Benchmark Is Not Enough
For many businesses, there isn't one benchmark. There are several.
A high-margin bundle may support a $90 CPA. An entry product may only support $25. A subscription offer may justify more upfront spend than a one-time purchase.
If your ads sell multiple product types but you evaluate performance against one blended CPA, you will make bad decisions. You may kill profitable campaigns. Or scale ones that quietly erode margin.
Meta cannot sort this out for you. When it reports your average CPA across a campaign, it is compressing fundamentally different economics into a single number.
The Problem With Averages
Most advertisers evaluate performance with averages: average CPA, average ROAS, average order value. Averages are convenient. They are also misleading.
Advertising performance is not a number. It's a distribution.
Here's what that means concretely.
Suppose your campaign reports an average CPA of $45. That sounds manageable. But when you look at the underlying data, you find that 30% of your conversions came at under $20, another 30% came in around $40, and the remaining 40% came at $70 or more.
You don't have a $45 CPA. You have a segment that's highly profitable, a segment that's marginal, and a segment that's losing money. The average hides all three.
The same applies to ROAS. A blended 3x ROAS can mask the fact that half your spend is generating 5x returns and the other half is generating 1.5x. Those require completely different responses — but if you only see the average, you'll treat them as one problem.
This is especially dangerous when scaling. As we described in Meta Is Not Sequencing Your Ads, the pacing system is designed to spend your budget. As budget increases, the system expands into lower-probability auctions. Marginal CPA rises. But the average moves slowly — it takes time for the growing tail of expensive conversions to visibly drag down the blended number. By the time the average looks bad, the problem has been compounding for days or weeks.
Without understanding how CPA distributes across segments, how marginal efficiency changes as spend increases, and whether scale is degrading conversion quality, you are reacting to surface metrics. With distributions, you manage risk.
The Real Risk Isn't Meta
Meta is doing exactly what it is designed to do. It maximizes predicted event probability. It spends the budget you give it. It adapts to auction conditions in real time.
It does not optimize for your margin. It does not optimize for your payback period. It does not optimize for your capital constraints.
And it can't.
The real risk isn't that the algorithm is broken. The real risk is assuming it understands economics it was never built to see.
A Simple Experiment
Pull your last 90 days of acquisition data.
Calculate your actual break-even CPA by product — not by campaign, by product. Include COGS, shipping, payment fees, and refunds.
Now compare that to the CPA you've been targeting.
If they're different — and for most advertisers, they are — you've been giving the system the wrong constraint. Every day the gap exists, the pacing system is faithfully spending your budget against a target that doesn't reflect your economics.
Closing that gap won't change how Meta's algorithm works. But it will change what you're asking it to do. And that's the one variable you actually control.