Marginal ROAS: when to stop spending.

What did your last $50,000 of ad spend buy?

Most brands can answer a different question with great precision: what did all of it buy, on average. The dashboard says 3.0. The monthly review says the account is healthy. And yet the business has less cash than it did at a lower budget, and nobody can point to the metric that went wrong.

Nothing went wrong with the metric. It was answering the wrong question. Average ROAS describes the account you have. Marginal ROAS describes the decision you are about to make, which is almost always whether to spend more or less than you spend now. This post is about estimating that second number honestly, and about the rule that tells you when to stop.

The short version

Keep adding budget while marginal ROAS is above breakeven ROAS, which is one divided by your contribution margin. Stop when it falls below. Because marginal ROAS is always an estimate, the rule works on the interval, not the point. Scale when the whole interval clears breakeven, hold when it straddles, and step back when the whole interval sits below it.

Everything else in this post is how to get that interval, and how to know which kind of number you are holding once you have it.

Average ROAS describes the past. Marginal ROAS describes the decision.

Average ROAS divides all revenue by all spend. Marginal ROAS divides the change in revenue by the change in spend between two budget levels. If you move from $100,000 to $150,000 a month and store revenue moves from $420,000 to $570,000, the marginal ROAS on that step is $150,000 divided by $50,000, or 3.0, whatever the average says.

The two numbers diverge because returns diminish. The first dollars in an account reach the most responsive people: those already searching for the brand, recent visitors, the audiences an algorithm finds cheapest. Each further increment has to reach someone less ready to buy. So each increment returns less than the one before it, and the average falls more slowly than the margin, because it is still carrying all the early dollars.

Google’s open-source mix modeling project describes the same idea in its documentation. Return on investment is the average over all spend, marginal return is the return on the next dollar, and when the marginal figure is much lower than the average the channel is saturating. It is the vocabulary a finance team uses for any other investment. Advertising is not special. It is just unusually good at reporting its average.

How a healthy average hides a loss

Here is an illustrative brand at four spend levels. The numbers are arithmetic, not a client result. Contribution margin is 40% after cost of goods, shipping, payment fees, and returns, so breakeven ROAS is 1 divided by 0.40, or 2.5. Revenue is total store revenue, including sales that would have happened anyway. Differencing it between rows isolates what the extra spend added, provided nothing else changed between rows.

Read down the average column and the account looks fine at every level. Even at $250,000 the average of 3.0 sits comfortably above the 2.5 breakeven, partly because it includes sales that would have happened anyway. Read down the marginal column and the story changes at the second step. The move from $150,000 to $200,000 returned 2.2 on the new money, below breakeven, and contribution fell by $6,000. The move to $250,000 returned 1.4, and contribution fell by another $22,000.

Profit peaked at $150,000. The last $100,000 of spend cost the business $28,000 of contribution while the average never dropped below breakeven. That is the whole problem in one table. The average can only warn you after the margin has been losing money for a while, because it is diluted by every profitable dollar that came before.

Why the stop line is breakeven, not your target

The stop line comes from contribution, not from a number someone set in a planning deck. A dollar of spend that returns R dollars of revenue produces R times the contribution margin in contribution. It pays for itself when that is at least one dollar, so the line sits where R equals one divided by contribution margin. At a 40% margin that is 2.5. At 60% it is about 1.67. The breakeven ROAS calculator does the arithmetic for your own margin, and the MER target calculator turns the same inputs into a store-level efficiency target.

Two refinements matter. First, the revenue in that formula has to be incremental revenue, the sales that would not have happened without the spend. Platform-attributed revenue is not that number, a gap MER vs ROAS walks through in detail. Second, the margin should be the variable margin on an incremental order, which is often lower than the blended margin because incremental orders skew toward discounted and first-time purchases. If you already bid to margin, profit on ad spend applies the same logic at the campaign level.

Three ways to estimate the next dollar, and how far off each can be

Marginal ROAS is never read off a report. It is estimated, and the method decides what kind of number you have. There are three practical methods for a brand without a dedicated data science team, plus one tempting shortcut. They differ less in difficulty than in what they can prove.

Method one: a deliberate spend step

Raise or cut spend by a planned amount, hold it long enough to see the response, and compare store revenue before and after. This is the table above done for real, one row at a time. It is a before-and-after comparison, so what it measures is correlation in time. Anything else that changed in those weeks, a promotion, a price change, a competitor’s stockout, lands in your estimate.

Its error bar is easy to compute and usually sobering. Suppose weekly store revenue varies with a standard deviation of $12,000 even when nothing changes, and you add $12,500 a week of spend, comparing four weeks before with four weeks after. The standard error of the difference in average weekly revenue is 12,000 times the square root of one quarter plus one quarter, about $8,485. Divided by the $12,500 step, the standard error on marginal ROAS is about 0.68. If the step returned a point estimate of 1.4, the 95% interval runs from roughly 0.07 to 2.73.

That interval straddles the 2.5 breakeven line, so the honest verdict is that you do not know yet. Double the step to $25,000 a week and the same noise produces a standard error of about 0.34. The same 1.4 point estimate now carries an interval of roughly 0.73 to 2.07, entirely below breakeven, and the verdict is clear. Step size buys precision in direct proportion, which is why timid tests so rarely conclude anything.

The trade-off is that a bigger step measures the average return across a wider stretch of the curve, the chord rather than the slope at your current spend. In practice that is acceptable. You are choosing between budget levels, not computing a derivative.

Method two: day-level spend against revenue

The second method uses history you already have. Plot daily spend against daily store revenue, fit a curve, and read the slope at your current spend. It costs nothing and feels scientific. It is also the weakest of the three, because the relationship it measures runs in both directions.

Budgets and bidding systems respond to demand. When more people are shopping, automated bidding finds more auctions worth entering and spend rises. When a promotion lands, the team raises budgets. So high-spend days are disproportionately high-demand days, and the fitted slope credits the ads with demand they were chasing. The bias usually runs upward, telling you the next dollar is worth more than it is. Treat this as correlation, useful for spotting the rough shape of diminishing returns, never as the number you bet the budget on.

Marketing mix models are the disciplined version of this method: weekly data, controls for seasonality and promotions, and explicit curves for carryover and saturation. Google’s research on Bayesian mix models computes marginal return directly from those curves, and is candid that with small samples the priors have a large effect on the estimates. The output is modeled contribution. It is far better than a scatter plot, and it is still not an experiment.

Method three: a holdout

The third method withholds spend from a randomly chosen or carefully matched part of the market, a set of regions or a slice of an audience, and compares outcomes against the part that kept spending. Because the comparison happens at the same time, seasonality and promotions hit both sides equally. This is experimentally measured incrementality, the only kind of number here that can claim cause.

The catch is which question a holdout answers. Switching a channel off in the test regions measures the average incremental return of the whole channel, not the marginal return of its last dollar. To measure the marginal you would hold spend at, say, 70% in the test regions rather than zero, which shrinks the effect you are trying to detect and makes the test harder to power. Incrementality testing without a data team sets out that power arithmetic, and it is the reason most brands run the average test first and reason about the marginal from it.

That reasoning is sound in one direction. Under diminishing returns the marginal sits below the average. So if a full holdout says a channel’s average incremental return is 2.1 against a breakeven of 2.5, the marginal is lower still, and you have learned enough to cut. If the average clears breakeven comfortably, you have learned less about the next dollar than it seems.

The shortcut: platform simulators

Google Ads offers bid, budget, and target simulators, and Performance Planner forecasts results at different budgets. They are useful for one thing: seeing how much more inventory the platform believes exists at a higher budget. Google describes the simulators as estimates of what results could have been over the previous seven days at different settings, and those results are conversions as the platform attributes them. A simulator can tell you the next dollar buys more attributed conversions. It cannot tell you whether any of them are incremental. Treat its curve as an upper bound and never as the input to the stop rule.

The stop rule, with error bars

Put the pieces together and the rule has three zones. Each is a decision you can write down before the data arrives, which is the easiest way to stop a team from arguing a result into the answer it wanted.

1. Scale when the lower end of the marginal ROAS interval is above breakeven. The next step is profitable even on a pessimistic reading, so take it and measure again.
2. Hold when the interval straddles breakeven. You do not know yet, so wait for more weeks of data or take a larger step to shrink the interval. Do not scale on a point estimate that sits inside the noise.
3. Step back when the upper end of the interval is below breakeven. Even the optimistic reading loses contribution. Return to the last level where the rule said scale.

Write down which method produced the interval. A 2.7 from a holdout, a 2.7 from a spend step, and a 2.7 from a platform simulator are three different claims: experimentally measured incrementality, a before-and-after correlation, and a platform-modeled attributed estimate. Only the first can be called causal. The second is usually good enough to steer by when nothing else changed. The third should never decide a budget.

What the common advice gets wrong

The common advice is to scale until ROAS reaches the target. It sounds disciplined. It is quietly a rule for spending past profit.

Look at how the platforms define a target. Google’s documentation for Target ROAS bidding says the system sets bids to maximize conversion value while trying to achieve an average return on ad spend equal to your target, and that some conversions will return more and some less. An average held at the target means the marginal auctions, the last ones the system chose to enter, return less than the target. Set the target at breakeven and the last dollars lose money by construction.

This is not a flaw in the bidding system. It is doing exactly what it was asked to do. The mistake is treating an average target as a stop line. If you want spend to stop near breakeven on the margin, the average target has to sit meaningfully above breakeven, and how far above depends on how steep your curve is, which is exactly what the three methods estimate.

The second thing the advice gets wrong is where it looks. Platform ROAS is attributed revenue divided by spend inside one platform. The stop rule needs incremental revenue measured at the store. A brand can watch platform ROAS hold steady while blended MER slides and store revenue barely moves, because the platform is claiming sales that were coming anyway.

When this advice is wrong

The breakeven stop rule assumes the only value of a sale is the contribution from that first order. Several real situations break the assumption. In each one the right move is to change the inputs, not to abandon the rule.

• →Customers who reliably come back. If new customers buy again, the value of acquiring one is contribution over your payback window, not first-order contribution. Use that figure and the breakeven line moves down. Be honest about the window, and about how much of the repeat revenue would have arrived anyway.
• →Inventory that has to move. When stock is aging, the alternative to selling it with ads is a deeper discount or a write-down, so the true margin on those units is higher than the accounting margin. The rule still applies, with the adjusted margin.
• →Fixed costs that arrive in steps. The rule is about variable contribution. If the next level of volume needs a new warehouse shift or another support hire, the step carries a fixed cost the per-order margin does not see.
• →Deliberate peak-season acquisition. Some brands choose to buy customers below breakeven in November because those customers repeat in the new year. That is a legitimate bet, as long as it is written down as a bet with a payback assumption rather than discovered afterward as a loss.
• →Effects slower than the test. Upper-funnel spend often works on a lag longer than a four-week step. A step that shows nothing may be measuring the wrong window rather than a dead channel.

What to do this month

It is November, the worst month to learn your curve and the most expensive month not to know it. Spend is at its highest, demand is at its most volatile, and every before-and-after comparison is contaminated by the calendar. Do not run a clean spend-step test through Black Friday. Do use what you already know.

1. Compute breakeven ROAS for your real contribution margin this month, including the deeper promotional discounts. A margin that drops from 45% to 35% under promotion moves breakeven from about 2.22 to about 2.86, and most accounts never update the line.
2. Find your last deliberate budget increase outside a peak period, if you have one, and compute the marginal ROAS on that step from store revenue. It is your best available estimate of where the curve sits.
3. Set peak budgets in steps with a stop rule attached, as in the week-by-week November budget plan, so a flat response triggers a cut rather than another increase.
4. If the margin itself is the constraint, work on it directly. Margin, offers, and order value often move the breakeven line further than any bidding change can.
5. Put a spend-step or holdout test in the January plan, when demand is steadier and the answer will be cleaner.

Most brands at this size are not short of data. They are short of a rule that turns the data into a budget decision. If you want an independent read of where your spend goes, start with the free audit. It is a read of your Google Ads or Meta account, and it is useful whether or not you ever work with us.