Average vs marginal iROAS: which one should you scale on?

Average iROAS tells you what a channel returned across the spend you already ran. Marginal iROAS estimates what a small increase from where you are now would return. If the decision is where the next dollar goes, marginal is the number closer to it.

But marginal iROAS is not an observation of the next dollar. It is a local reading off a modeled response curve, and that curve assumes things: how the added spend gets deployed, what else was moving, and what happens at spend levels your data has never seen. Treating that reading as a hard fact is where budgets quietly break. That is a bigger question than average versus marginal, and it is the one your budget actually turns on.

What's the difference between average and marginal iROAS?

Average iROAS is a channel's total incremental revenue divided by its total spend, a reporting number for how the channel did overall. Marginal iROAS is the return the model expects from a small increase near your current spend, an allocation number for whether the next dollar is worth adding. A channel can look strong on average and be weak at the margin.

They diverge because of saturation. Early dollars reach the cheapest, highest-intent audience; later ones reach colder, more expensive attention and return less. Efficient early spend keeps the average high long after the return at the edge has fallen. That is the common case. Some curves are S-shaped and start slow before they catch.

Say a channel produced $300,000 on $100,000 of spend. Average iROAS is 3.0x, so it looks like your best performer. But if the curve says a small increase from here returns 1.4x, the next dollar is a 1.4x dollar. You are reading the average and buying the margin.

How is marginal iROAS calculated?

Every MMM reads it off the response curve of spend against incremental revenue: the marginal number is the slope of that curve where you sit. That matters for how far to trust it. Marginal iROAS is not a measured return on a dollar you already spent; it is the slope of a curve the model estimated from your data, so it carries that curve's uncertainty and its assumptions. The method varies. Some take the analytic slope; Google's Meridian uses the return on a small finite step above historical spend. Either way the estimand is the same, the return on the next increment.

The number quietly assumes you spend the next dollar the way you spent the last ones. Meridian's curves hold your historical flighting fixed and assume cost per unit of media stays at its historical average. Change your CPMs, geo mix, audience, or creative, and the realized return drifts from the quoted one. So "the next dollar returns 2.7x" honestly reads: under this model, and if you deploy the spend as you have been, about 2.7x. That is the version you can defend to a skeptical CFO.

The marginal number also usually carries a wider range than the total the same model reports, because a slope is more sensitive to wobble in the curve's shape than its level. That is a reason not to print it naked, as one confident figure. A Bayesian model has the full distribution and can hand you the range. Most reports still print one decimal.

Does your data actually cover the spend you're asking about?

Start with what "covered" has to mean. Not the spend levels you have operated at, but the levels your data actually varies across, enough to show the slope of the curve there. If your spend sat flat near one level all year, the model has little to learn that slope from, even though you were technically in range the whole time. Where the data carries that variation, the curve is anchored. Push past it and the model is guessing at a curve it has never seen, a risk Meridian says grows the further out you go. It bites hardest when you scale, because "add another 40%" usually asks about spend nothing in your history covers.

Do not oversell "covered" either, including us. Even inside that range, MMM runs on observational data, so if spend moved with demand, promotions, or seasonality, the curve can be biased however much data surrounds the point. Two separate questions live here. Does the data support this part of the curve, and do you have reason to read it causally at all. A spend level can pass the first and fail the second.

You settle it by creating the variation on purpose. If the model says moving from $100,000 to $140,000 adds $80,000, run a geo holdout that makes that contrast and see what comes back. Predicting held-out weeks only shows the model forecasts well; the experiment is what speaks to cause, and even then only for the contrast it created. It tells you about the return in that band, not the slope at $200,000. Proving a channel works near today's spend is a different thing from mapping the shape of the whole curve, and treating them as one is how a single result gets stretched across a range it never covered.

Does the marginal number clear your breakeven?

Marginal iROAS is revenue, not profit. A 1.0x return loses money at any margin under 100%, so breakeven is not 1.0x. It is 1 divided by your contribution margin, so 2.5x at a 40% margin and 5.0x at 20%. Use the margin on the revenue the new spend actually creates, since returns, discounts, and new-customer economics can make incremental dollars thinner than your blended number.

The test is marginal iROAS times your incremental margin, minus one. Above zero the next dollar adds profit; below zero it destroys it. Clearing that line is necessary, not sufficient, and where you set the bar is a risk policy you choose, not a statistical law. The cleaner question, from a Bayesian model, is how likely the true return is to clear breakeven and how much certainty you demand. Finance might want 90% before adding budget; a business long on inventory might move at 70%. The model quantifies the uncertainty. You decide how much to buy.

Should you scale, or just reallocate?

Breakeven answers whether to grow the total budget. It cannot tell you where a fixed next dollar should go, and that is the question most explainers skip.

Say Meta shows 3.0x marginal iROAS and Google shows 4.2x, at similar margin and risk. Meta clears breakeven comfortably, and the next dollar still belongs in Google, because that is where it works hardest. You move budget toward the higher marginal return until the two converge or you hit a constraint. The question that matters is where the next dollar works hardest across every channel you could put it in. A profitable channel can still be the wrong one to feed.

Which iROAS should you actually scale on?

Marginal, for marginal decisions, but never the point estimate alone, because the math is worthless if the curve underneath is not causal. Platform-reported ROAS is attribution: it credits conversions near an ad without telling you what would have happened otherwise, and a fancier observational model does not fix that. Gordon, Zettelmeyer, Bhargava, and Chapsky tested observational methods against 15 randomized experiments at Facebook and found they frequently failed to recover the experimental result even after heavy adjustment. Selection bias runs deep enough that careful adjustment does not reliably recover the causal number, and the miss can go either way.

An MMM does not escape this by fitting a prettier curve. It earns a causal reading only when its assumptions hold, and experiments are what shore them up. At Stella the response curve is calibrated with your own holdout results, so the effect is disciplined by real experiments, not correlation alone. Calibration ties the curve to a measured contrast; it does not hand you the whole shape, which still leans on the spend variation in your data and the form you assumed. That makes the marginal number more defensible, not certain: an experiment in other markets or another period may not measure exactly what you are scaling, and turning it into a prior adds its own uncertainty. That is the bar we hold our own model to, and it is how we grade our own work. The holdout that calibrates the model is where the causal signal comes from.

So before you move money, ask what the number estimates and how the spend gets deployed, whether it sits inside your data or past it, what makes the curve causal, and whether it beats both your hurdle and your next-best alternative. A marginal iROAS that arrives with the curve, the range, and the experiment behind it is a number you can act on. A single confident decimal is not a strategy. It is a decimal with a recommendation attached.

FAQ

What is marginal iROAS?

The incremental revenue the model expects from a small increase in spend near your current level, read as the local slope of a channel's response curve or a small step along it. It tells you whether the next dollar is still worth adding.

Can a channel have high average iROAS but low marginal iROAS?

Yes, and under diminishing returns it is common. Efficient early spend holds the average up after the return on additional spend has fallen. The exact shape is model- and channel-dependent, so treat it as the usual case rather than a rule.

What breakeven should I compare marginal iROAS against?

Your breakeven multiple, 1 divided by your incremental contribution margin, not 1.0x. A 40% margin implies 2.5x. Use the economics of the revenue the new spend creates, and include other marginal costs when they matter.

Should I scale whenever marginal iROAS clears breakeven?

Not necessarily. Clearing breakeven means the next dollar is probably profitable. It does not mean it is the best use of that dollar. If another channel has a higher risk-adjusted marginal return, the money belongs there first.

How do I know if my vendor's marginal iROAS is trustworthy?

Ask what it estimates, what spend range supports it, what makes the curve causal, what experiment informed it, and how uncertain it is. If the answer is one decimal and nothing else, you cannot move a budget on it.

Average iROAS tells you what happened. Marginal iROAS is closer to the decision in front of you. But the number your budget should move on is the risk-adjusted marginal contribution, off a curve you have reason to believe, against the next-best home for the dollar. That is a higher bar than an mROI in a dashboard, and it is the one worth holding every vendor to, us included. Book a demo and we will show you where your curve is guessing, how wide the range really is, and what we tested to trust it.