We ran every defensible version of the analysis on the same eight months of data. The answers ranged from below zero to 14x. The default settings landed at the flattering end.
Someone asks me about marginal ROAS every few weeks, so we went and pulled the number apart.
We took eight months of one brand's paid search. Real spend, real revenue, properly attributed. Then instead of picking one analysis and reporting it, we ran the lot. Daily, weekly, monthly buckets. Different start dates. Different attribution models. With and without a carryover adjustment. With and without separating ad-driven sales from the demand that was already there. Around 1,800 combinations, every one of them a reasonable thing for an analyst to do.
The answers ran from a marginal ROAS below zero to about 14x.
Same data. Same channel. Same eight months. The only thing that changed was the choices an analyst is free to make, and those choices moved the answer from "this is torching money" to "this is the best thing we own".
Then we checked where the defaults land. Most of these tools ship with one particular set of those choices baked in, and that combination sat at the 95th percentile of everything we generated. Roughly a quarter of the perfectly reasonable analyses had the channel underwater at the margin. You would never see them, because the tool shows you a single number and it happens to be one of the rosiest in the room.
The pitch is half right
I want to be fair to the argument, because half of it is sound.
Average ROAS does blend your brilliant spend and your wasted spend into one comfortable figure. That figure will happily tell you a channel is doing 4x long after the last slice of budget stopped paying for itself. Diminishing returns are real and average ROAS hides them. Fine. Agreed.
Where I get off the bus is the second half: the assumption that the marginal number is sitting in your history, waiting to be found.
Marginal ROAS is a slope. It's the steepness of your spend-to-revenue curve at the point you're spending now, and knowing a slope means changing the level of spend and watching how revenue responds. Nothing in your historical data was changed on purpose. What the tool has done instead is chop your history into periods, line spend up against revenue in each bucket, and read the gradient off that.
Somebody chose the period before any measuring happened.
The sequence is fictional too. You don't spend pound one, watch it work, then decide on pound two. Budgets go out at once into a continuous auction and the revenue dribbles back over the following weeks. "Your 11th pound returned 1.3x" is a point on a fitted curve wearing the costume of a diary entry.
Three things the window is quietly doing
- Where you cut. A day, a week and a month give you three different slopes. None of them is the true one.
- Carryover. A pound you spend today keeps working for days or weeks, so this week's revenue is a smear of the last month's spend. Line them up week against week and you misattribute the lot.
- Demand. You spend more in exactly the weeks demand is already high. Pre-season, payday, the sale, the cold snap. Spend and sales rise together whether the advertising did anything or not, and the model sees the two moving in step and credits the ads.
The last one is the big one, and it's the one the marketing never mentions. In our 1,800 runs, the lever everyone sells you, the clever response curve fitted to your data, barely moved the answer. The two things that genuinely moved it were carryover and the demand adjustment, which are the two the defaults skip. The dial that gets the sales pitch is the dial that doesn't matter.
The same method lies in the opposite direction on social
We ran the identical process on the same brand's social prospecting and got the mirror image.
Most of the analyses put it below break-even. The curve buried it, because last-click attribution simply can't see the work upper-funnel activity does. Same method, same brand, same eight months, now under-selling a channel as confidently as it over-sold the other one.
This is what should worry you. The method doesn't hand you a noisy number that's equally likely to be wrong in either direction. The direction it lies in depends on the channel. It flatters the channels that harvest demand you already had, and it buries the ones that create demand you didn't yet have. Optimise on it without thinking and you'll pour money into capturing existing intent while starving the things that build it.
A glimpse by accident
One brand we work with had its paid search switched off for a few days. Not an experiment, just one of those things. But it left a fingerprint.
The moment the ads went dark, organic search revenue jumped. It roughly tripled, and it stayed up for as long as paid was off. The ads came back on and organic dropped straight back to where it had been. The two moved like a see-saw.
Doing the sums on the cleanest part of that window, something like 85% of the revenue paid search had been claiming reappeared in organic the instant paid switched off. The customers were still coming and still buying. They were clicking the free link instead of the paid one. On a like-for-like basis the genuinely incremental return looked much closer to break-even than to the 8x the dashboard was reporting.
I'll be straight about the quality of that evidence. It was a handful of days, tangled up with a promotion ending at roughly the same time. And a few days is not proof and I wouldn't sell it as such.
What makes it interesting is that a messy accident and a careful model, two methods that share no assumptions, pointed at the same answer: the headline number was mostly demand the brand already had. That's the moment you stop arguing about it on a slide and go and run the test properly.
Earn the number
So do we bin the idea? No. Diminishing returns are real and you should absolutely spend less where the next pound/dollar/euro is weak. But you have to earn that number rather than lift it off your history and call it a fact.
Earning it means breaking the link between spend and demand on purpose. Try a holdout or a geo test, or a deliberate step up or down in spend. Then the change you measure is yours rather than the market moving underneath you. Let the data tell you how long a pound keeps working instead of assuming. And show your working: which window, which assumptions, how confident, with an honest flag on every channel telling you whether this is a measured effect or a hopeful one.
But back to that paid search channel. The number a tool would hand you was around 8x. Strip out the demand it would have captured anyway, allow for a pound working over weeks rather than landing all at once, and the best modelled estimate sits near break-even. Call it 0.7x, with error bars wide enough to cross 1 in both directions.
That's what we'd put on the page, alongside a flag reading "modelled, not yet tested" and a recommendation: if this number is about to move real budget, run a geo holdout and find out for certain rather than betting the quarter on a slope we fitted to last winter.
A marginal ROAS handed over with no window, no assumptions and no uncertainty is a decision somebody else made, dressed up as a fact. I'd rather give you an honest range than a confident point that turns out to be wrong. That's how we build things at Ringside: tell you what we know, flag what we don't, and never sell a choice as a discovery.
It's a less exciting slide, and it's true.