A campaign can show a 4.0 ROAS in the ad platform (four dollars of revenue per dollar of ads) and still lose money on every extra dollar you add. The gap has three causes. Each new customer costs more as spend grows. The platform counts sales the ads did not cause. And teams judge groups of new customers before they have shown how much they will really spend.
Ad platform reporting is built to make the platform look good. That alone is a reason for product managers and growth teams to keep their own economic model and run every budget decision through it. What follows is that model: how to price CAC, where ROAS and MER mislead, how lift is measured, which attribution errors recur, and how cohorts decide whether spend can scale.
Marginal CAC decides where paid spend saturates
Customer Acquisition Cost (CAC) comes in two versions, and they answer different questions. Blended CAC = total spend ÷ total customers acquired. It hides inefficiency and overstates how far a channel can scale. Marginal CAC = the cost of the next customer. It shows where a channel saturates.
A small example shows the difference. A channel spends $20,000 a month and brings 400 customers: blended CAC is $50. The team raises spend to $30,000 and gets 520 customers. Blended CAC is now about $58, which looks like mild inflation. The extra $10,000, though, bought only 120 customers. Their marginal CAC is $10,000 ÷ 120, about $83, two thirds above the old average. If the business can afford $70 per customer, the increase already loses money, and the blended number hides it.
So teams model the whole curve: CAC at each budget step, how fast it bends as spend rises, and the worst case next to the best case. That keeps the next dollar in view instead of the average one.
True CAC also includes more than media. Creative production, marketing tooling, data fees, experiment overhead and attribution infrastructure all belong in it. Leave them out and every ratio built on CAC looks better than it is.
CAC only means something next to LTV, the money a customer brings over their whole lifetime. Common working limits are CAC under ⅓ of LTV for consumer models, under 20-30% of LTV for SaaS, and under ⅕ of supply LTV for marketplaces. Without a limit like this, budgets keep growing, because nothing in the platform dashboard pushes back.
ROAS and MER both flatter paid spend
ROAS (Return on Ad Spend) is revenue ÷ spend. It is useful for small tactical calls: comparing campaigns, ranking creative, reading audience tests. It breaks down as soon as real economics come in. It cannot tell which sales the ads caused. It should not decide budgets across channels. It misreads sales that come long after the click. And it flatters subscription models, where revenue arrives months later.
MER (Marketing Efficiency Ratio) is total revenue ÷ total marketing spend, sometimes called "blended ROAS". Its strength is that it ignores the fight over which channel gets credit. It ties marketing to real revenue and smooths out noisy channels by looking at the whole picture.
On its own, MER has the opposite blind spot. It cannot separate sales that came from ads from sales that would have come anyway. It hides the point where each extra dollar starts buying less. A burst of early revenue from new customers can push it up or down. So use MER as a top-level health check on the total budget, and make channel decisions elsewhere.
The numbers show why. Say total revenue is $1.2 million on $300,000 of marketing: MER is 4.0. If $700,000 of that revenue would have arrived without any paid media, the paid-driven part is $500,000. The effective return on marketing is then about 1.7, less than half the headline.
Both ratios mislead in the same underlying way. Platforms take credit for organic conversions, brand-driven demand, affiliate influence and retargeted buyers who would have converted anyway. Only incremental lift separates what the spend caused from what it merely touched.
Lift is the only figure that proves the spend worked
Incrementality, or lift, measures what the spend actually caused: the difference between what happened and what would have happened without it. Each test design covers a different slice and gives a different level of precision.
- Geo lift tests randomly assign markets or regions to higher spend. They suit large budgets, capture cross-channel effects and keep attribution bias low.
- Conversion lift tests compare treated and control users at the user level.
- Channel on/off tests switch a paid channel off for a short window to see what survives.
- Audience split tests compare exposed and unexposed cohorts.
Trust any of them only after checking two things: the sample was big enough, and the result is unlikely to be chance. Stopping early is the classic trap. If you check a lift test every morning and stop on the first good-looking day, you will see false wins far more often than the 5% you planned for. A lift number without these checks is noise with a decimal point.
How far off can the alternatives be? Gordon, Moakler and Zettelmeyer (Marketing Science, 2022) compared 663 large Facebook ad experiments with non-experimental estimates of the same campaigns. They concluded that those methods could not reliably estimate what a campaign caused. The median lift from their best statistical models came out at 83%, 58% and 24% for upper, mid and lower-funnel campaigns, against 29%, 18% and 5% from the randomized tests. The observational numbers overstated lift at every stage of the funnel.
A geo holdout in five decisions
- Unit and split. Use regions that buy independently (cities or metro areas, not neighboring postcodes that share commuters) and assign them at random, stratified by past revenue so both groups start alike. With only a handful of regions, pair them by past sales and randomize inside each pair.
- Holdout size. A smaller holdout is cheaper and noisier. Decide the share from the smallest lift worth detecting, then accept the revenue the dark regions will forgo.
- Pre-period. Record several weeks of sales in both groups before the change, so the readout compares how each group moved rather than raw levels.
- Duration. At least one full purchase cycle plus the usual conversion lag, fixed before launch, with no early stop on a good week.
- Readout. The extra conversions are those in the test regions minus what the holdout regions' trend predicts. Put the range of likely values in the budget memo next to the single best estimate.
A lift above 0 means the paid spend has a real effect. A lift of zero or below does not prove on its own that the ads only took sales that would have happened anyway. Read it together with its margin of error and the test design. Applied to the core metrics:
Incremental CAC = spend ÷ incremental conversions; incremental ROAS = incremental revenue ÷ spend.
An incremental CAC above LTV is unscalable. An incremental ROAS below 1 destroys value. A worked example makes the gap concrete (illustrative figures, not benchmarks):
- A campaign spends $50,000 and the platform attributes $200,000 in revenue → reported ROAS = 200,000 ÷ 50,000 = 4.0.
- A geo holdout shows only 40% of those conversions were incremental; the rest would have arrived through organic and branded demand, and here we assume the incremental orders carry the same average revenue as the rest. Incremental revenue = $80,000 → incremental ROAS = 80,000 ÷ 50,000 = 1.6.
- At a 50% contribution margin, incremental gross profit = $40,000, below the $50,000 spent → the campaign destroys value on the marginal dollar despite a headline 4.0 ROAS.
The same arithmetic works as a calculator for any channel:
Contribution per $1 of spend = reported ROAS × incremental share × contribution margin, minus the $1 spent
For the campaign above, 4.0 × 0.4 × 0.5 = 0.80, so each dollar returns 80 cents of contribution and loses 20. Turned around, the break-even reported ROAS is 1 ÷ (incremental share × margin). At a 40% incremental share and 50% margin that is 5.0. A prospecting channel with a 60% incremental share breaks even at about 3.3; a retargeting campaign where only 25% of conversions are incremental needs a reported ROAS of 8.0 just to pay for itself.
The reported number and the causal number point in opposite directions. That is why the holdout decides whether to scale.
Paid channels take credit for organic sales
Attribution systems produce that gap routinely, because they read correlation as causation. The errors repeat. Retargeting claims bottom-funnel buyers who were already coming. Branded search mostly harvests brand equity that existed before the ad. Multi-touch models quietly overweight the last click. Platforms report their own conversions generously. And any campaign aimed at high-intent segments looks like a winner. Each error over-credits paid, and over-crediting leads straight to overspending.
The defense is to check one method against another. Marketing mix modeling uses months of sales and spend data to estimate long-term impact. Google's open-source model, Meridian, generally available since January 2025, can start from the results of your lift tests, which is how the two methods should work together. Multi-touch attribution shows individual user paths. Lift tests give the ground truth. Scenario models check whether the conclusion holds under different assumptions. Simple questions close the loop, for example a post-purchase survey asking buyers where they first heard of the product. If the survey says 30% came through a podcast that attribution credits with 3%, one of the two is wrong, and the lift test decides which.
Used this way, attribution stops being a report and starts steering the decisions that spend money: budget allocation, creative strategy, new channels and the marginal CAC limit for each channel.
Platform ROAS does have one thing going for it: speed. Lift tests are slow and they cost money. Holding out 10% of regions on a $500,000 monthly budget means forgoing whatever those regions would have bought from $50,000 of ads for the length of the test. For small budgets that cost is real, and directional ROAS may be the only practical signal. The cost argument weakens as spend grows. Once a channel takes six figures a month, one wrong scaling decision costs more than the holdout.
Cohorts gate the budget
Attribution tells you what a customer cost. Whether that customer was worth buying is a cohort question. Paid acquisition pays off only when new users form profitable cohorts.
The signals to watch are the shape of those cohorts over time: retention curves, cohort LTV, CAC payback, churn, engagement depth and, in B2B and SaaS, expansion revenue. Together they show whether budget can grow without blowing up burn.
Early data is the trap. It almost always overstates retention, ARPU, cross-sell and payback, so a team should not scale on cohorts that have not stabilized. Zumsteg's widely shared essay "Unchecked AB Testing Destroys Everything it Touches" (2022) makes the same point from the product side: a 10% first-order discount can win a short test and still erode the customer base the business depends on.
Cohorts work best as a gate on spend. Spend goes up when new cohorts show rising LTV, stable retention, faster payback and falling churn. All four together mean the engine is compounding. Spend comes down when CAC climbs faster than LTV, cohorts degrade, activation falls or competitive costs rise. Any one of those means the next dollar buys less than the last.
The budget model rests on three numbers: what a customer brings over their lifetime (LTV), what they cost (CAC) and how fast they pay that cost back. A healthy channel brings at least 3 dollars of LTV for every dollar of CAC. For SaaS, where CAC usually sits at 20-30% of LTV, that means about 3.3 to 5. Payback should be fast and CAC predictable. As rough ranges, payback under 6-8 months in B2C and under 12-18 months in B2B SaaS keeps cash burn manageable.
Then test how fragile those ratios are. What happens to payback if churn rises, or if CAC climbs faster than planned? A unit-economics model holds the arithmetic; a significance calculator and a scenario planner cover the test and the forecast.
Budget allocation should then follow marginal economics. The next dollar goes where incremental ROAS is highest, marginal CAC lowest, payback fastest and cohorts strongest, in that order. Volatile channels also need scenarios: CPC and CPM inflation, platform algorithm changes, competitive shocks, creative fatigue and geo expansion. Scenario work prevents overexposure before it becomes a write-off.
Measure lift before you argue about attribution
The method breaks wherever the team has a gap. PMs, growth leads and analysts all need to understand the funnel, marginal CAC, lift tests, attribution checks and cohorts. A fixed rhythm keeps it honest: channel reviews every week, CAC/LTV checks every month, a fresh look at scenarios every quarter.
Before the next budget increase, run three checks in this order. First, get an incremental ROAS from a geo or conversion lift test on the channel you want to scale, run for at least one full purchase cycle. If the extra gross profit does not cover spend, as in the worked example, there is nothing left to argue about in attribution. Second, price the proposed increase at marginal CAC and hold it against the CAC/LTV boundary for your model. Third, confirm that the latest cohorts have stabilized before trusting their LTV.
A channel gets more money only after passing all three. MER answers a separate question: whether the portfolio as a whole still holds together after the shift.