“We use a multiplicative MMM model—because we believe everything in the world interacts.”
A vendor said this proudly to a room full of marketers, as if it were automatically more realistic.
It sounds realistic. Of course the world has interactions.
But “multiplicative” isn’t just a technical preference. It’s a structural commitment—a decision about what kinds of stories the model is allowed to tell.
And some of those stories are not merely aggressive. They are operationally impossible.
An additive model says channels contribute amounts that sum. A multiplicative model says channels scale outcomes—and sometimes each other. That structure choice determines what kinds of counterfactuals the model is even allowed to generate.
Additive models tend to produce stable narratives and stable ROI. Multiplicative models tend to produce conditional narratives—ROI that changes sharply with context. The question isn’t which is more advanced. It’s which kind of error is safer for your business.
When multiplicative can be defensible
If you can articulate a mechanism (not a vibe) and validate that co-activation truly amplifies outcomes—especially under controlled tests—multiplicative structure can be reasonable.
The problem is not interaction. The problem is assuming universal interaction by default.
What the model is really assuming
In most MMM contexts, “multiplicative” doesn’t mean “we allow a few interactions.”
It means the model is built so that channels don’t merely add contribution. They scale each other.
In that structure, increasing spend in channel A implicitly raises the effectiveness of other channels—because the model couples them.
Sometimes that’s defensible. Often it isn’t. And it creates two quiet failure modes.
Failure mode 1: The model forces interactions where no interaction exists
If you run TikTok to reach Gen Z and direct mail to reach an older segment, what is the mechanism that makes these two channels “multiply” each other?
Sometimes teams answer with a vague idea—brand halo, awareness spillover, general lift.
But this is where the structure becomes dangerous: you don’t need to prove the interaction. The model has already assumed it.
So you get a world where everything helps everything else—even channels that operate on different segments, different contexts, or different stages of the journey.
This isn’t “more realistic.”
It’s a prior on universal synergy disguised as a modeling choice.
Failure mode 2: Multiplicative structure tends to allow synergy, but not cancellation
Real interactions are not always positive.
Sometimes channels overlap and crowd each other out:
- two tactics hit the same audience
- two promotions cannibalize timing
- two demand-capture channels compete for the same last-click traffic
Those are cancellation interactions: the combined effect is less than the sum of individual effects.
Many multiplicative MMM structures, in practice, are better at expressing synergy than cancellation. They tend to assume that if both channels are “on,” the total effect scales up.
A segmentation sanity check
Imagine a market where two channels largely hit the same people.
A local newspaper campaign converts about 750 customers in a period.
Billboards also convert about 750 customers in the same period.
If the audience overlaps heavily, running both does not convert 1,500 customers. It often converts less than the sum, because you are paying twice to reach the same decision-makers.
An additive model can represent that overlap explicitly—by adding a targeted interaction term that can be negative (cancellation) when channels crowd each other out.
A multiplicative structure, by contrast, can make “both on” look like automatic uplift. That is how you get counterfactuals that imply you can convert more people than exist in the reachable segment.
One more version of the same problem shows up as frequency.
When the same buyer sees your message twelve times across two channels, the eleventh and twelfth exposures rarely “multiply” the first. They often do the opposite: they waste budget, create fatigue, or shift attention from persuasion to annoyance.
If your model assumes co-activation is automatically uplifting, it can mistake rising frequency for rising effectiveness—especially in periods when media spend increases broadly and audiences are saturated.
The most practical risk: ROI inflation via cross-channel scaling
Here’s the uncomfortable version of what multiplicative structure can do:
If one channel is currently not saturated and you increase spend there, the model can increase total marketing contribution in a way that looks disproportionate—because the incremental spend doesn’t just add impact; it lifts other channels too.
So you can see narratives like:
- “We increased spend in one channel, and overall marketing effectiveness doubled.”
Sometimes a big shift really happens.
But if the mechanism is not defensible, this is not insight. It’s structural leverage—a model giving you the story it was built to allow.
And because budgets respond to ROI rankings, this is not theoretical. It’s a spend decision.
When interaction is real, don’t bake it everywhere
None of this says interactions don’t exist.
Some are obvious:
- an incentive (cashback, coupon) working with a targeted push notification
- coordinated creative across two channels timed to the same offer
- retail media + in-store promotion
But the clean way to reflect this is usually selective interaction, not universal multiplication.
In an additive framework, you can model a specific interaction explicitly—one or two that you actually believe—and treat it like a hypothesis that must earn its place.
Not a default assumption that everything synergizes with everything.
A simple diagnostic: does the decision survive the structure?
One vendor I worked with tried to replicate their multiplicative results using an additive version of the same model.
They couldn’t.
The numbers were materially different.
That doesn’t prove the multiplicative model is wrong. It proves something more important:
the decision was not structurally stable.
If the ROI story flips when you change the model’s structure, you don’t have a measurement result.
You have a sensitivity problem.
And if you are allocating millions based on that story, the question isn’t:
“Which model is more advanced?”
It’s:
Which one is more decision-safe for this business?
Multiplicative MMM can be defensible in narrow cases, when you can explain the mechanism and validate the behavior under stress.
But “everything interacts” is not a mechanism. It is a sales line.
And when structure decides what’s possible, results stop being discoveries. They become permissions.