By Bradley Keefer, Chief Revenue Officer, Keen Decision Systems
The frustration with marketing mix modeling is real, and it is growing. Media sellers are exhausted by it, practitioners find it slow and easy to manipulate, and it is holding television back from operating at the pace of digital. These are fair observations but the industry just keeps arriving at the wrong conclusion.
The debate frames MMM as the problem. The model is too slow, too blunt, too easy to steer toward a convenient story. Fix the model, or stop relying on it so much. That is the conversation, but it is the wrong one.
The real problem is not what the model produces. It is what happens after.
Not all MMMs are the same
Before diagnosing what is broken, it helps to recognize that the industry is not actually talking about one thing when it says MMM.
The first is legacy MMM. Six-month engagements, six-figure consulting fees, a 90-slide deck delivered at the end of a fiscal quarter that nobody fully acts on before the next planning cycle begins. The speed critique is entirely fair here. When it takes months to produce a result, the decisions you can make with an outdated plan are inherently limited. This is where analysis paralysis lives, and it is a real problem.
The second is open-source and light-model tools. Platforms built on frameworks like Meta’s Robyn or Google’s Meridian were designed to break open the black box and accelerate the measurement process. They succeeded at transparency and speed. What they did not solve is model integrity. Speed without a well-calibrated, priors-informed model is not the win people were hoping for. A fast answer built on a hollow foundation is still a wrong answer, just delivered more efficiently.
The third camp is where other models sit. These tools set out to understand what objections come after speed. What emerges is the more interesting problem: even when you solve for speed, people still do not change their behavior as much as you would expect. That finding reframed everything about how we think about what MMM is actually for.
Where the critics are right
When MMM does what leadership wants it to say, it is not a modeling failure. It is a human one.
There is actually a scientific name for it. Bayesian reasoning, the same statistical framework that underpins modern MMM, starts from the idea that everyone brings priors into any analysis: beliefs formed from past experience that shape how new evidence gets interpreted. In a well-functioning model, those priors get updated as evidence accumulates. In a well-functioning organization, the same thing should happen. But organizations are not statistical models. When the evidence conflicts with what someone already believes, and when no structure exists to force a reckoning with that conflict, people do what people do. They find what they want to find. They discount what challenges the existing plan. The prior wins. The model loses. That is not cynicism. It is how humans process information under uncertainty, and no faster model solves it.
What solves it is structure. When a measurement output arrives as a report with no recommended action attached, no named owner, no decision deadline, and no reconciliation loop to close, the existing strategy wins by default. That is not the model’s fault. That is what happens when you hand someone a diagnosis without a treatment plan and then wonder why the patient did not change their behavior.
The speed critique is fair when applied to legacy MMM. But the premise that speed is the core problem does not hold once you have actually solved for speed. We know this from experience. The bottleneck was never the model runtime. It was everything the organization needed to do differently once the model finished running.
Where the conversation misses the point
A Harvard Business Review survey of global marketers found that 87 percent say MMM is important for data-driven decisions. Only 28 percent say their organization effectively converts those insights into timely action. That 59-point gap is not a measurement problem. It is a behavior change problem.
Most brands do all the work to get an MMM and then do not meaningfully change anything. They run the model, review the output, nod at the recommendations, and continue executing the plan they already built. When results do not improve, they question the model’s believability rather than their own compliance with what it recommended. The model becomes the scapegoat for a decision-making process that was never actually connected to it. They wanted the MMM to validate the strategy they already had, not challenge it.
MMM gives the broadest view across the full budget when no other tool can. That is accurate. But the point is not that MMM filled a gap. It is that measurement alone was never designed to close it. MMM was never meant to be the measurement system, the decision system, and the accountability system simultaneously. It was a measurement tool. The decision science layer, the forecasting layer, the reconciliation layer: those were supposed to exist around it. For most organizations, they do not.
That is exactly why closed-loop systems where measurement connects to a forward plan, the plan connects to a forecast with probability attached, and the forecast gets reconciled against what actually happened are a better alternative. Not better reporting. Faster learning. A system designed to show you how to fail faster, course-correct sooner, and make smarter decisions about tomorrow without abandoning the holistic view that makes MMM valuable in the first place.
Finally, MMM belongs at the cross-channel allocation level, not campaign optimization. However, this distinction only matters once organizations are actually acting on the macroeconomic output, and most are not.
Across the 400-plus brands in our data, flighted marginal ROI for media falls below one dollar, meaning brands are spending past the point of profitable return in the windows they are already running. When timing and allocation are optimized, that marginal ROI jumps to $5.84. Linear TV moves from $1.26 flighted to $6.64 optimized. That gap is not a modeling failure. It is evidence that the macro recommendation was seen and ignored.
Before the industry debates whether MMM should go deeper into campaign-level decisions, it should answer why it is not acting on the portfolio-level decisions it already has.
The three problems nobody is naming
The first is goals and incentives. If individual goals do not roll up to a shared top-line strategy, you cannot expect people to change their behavior after a model run. You are literally paying them not to. An agency briefed to deliver 500,000 impressions is not going to reallocate toward reach-light, revenue-heavy channels, even when the MMM says to, because their performance review does not reward incremental revenue. It rewards impressions. A brand team tasked with raising awareness five points will optimize for awareness, even if the model shows that reaching the wrong audience makes that awareness commercially worthless. Reach does not equal revenue. Awareness does not equal sales. Correlation is not causation, and no model fixes an incentive structure pointed in the wrong direction.
The second is data architecture. Everyone is quick to blame the MMM when output takes weeks. The model is rarely the bottleneck. The pipeline is. Inconsistent tagging, siloed sources, multi-week delays in ingesting actuals, no clean layer reconciling what was planned against what was actually spent. These are data governance failures, not measurement failures. The eMarketer MMM Trends report from earlier this year confirmed that data quality and integration are the top barriers marketers cite when converting insights into action, ahead of technical expertise, ahead of model complexity. Fix the pipeline and the model runs faster and tells you something you can use sooner. Blame the model and the pipeline stays broken.
The third is accountability. When a model produces a recommendation and the team does not follow it, who owns that gap? In most organizations, nobody. There is no reconciliation cadence comparing what was recommended to what was executed. There is no mechanism for explaining the variance. Without that loop, every subsequent model run starts from scratch. The organization does not learn. The model does not compound. And when results underperform, the model takes the blame for a decision that was never made.
This is the uncomfortable truth sitting underneath the entire MMM debate. Every organization says it is data-driven. The reconciliation data says otherwise. Being data-driven is not a philosophy or a hiring slide. It is a behavior, and behavior requires an incentive structure that rewards acting on what the data says even when it conflicts with what you already believed.
Most organizations have neither. They have dashboards, they have models, and they have meetings where people nod at recommendations that they have no intention of following. The data does not drive anything in that environment. It decorates the decision that was already made.
Until someone owns the gap between what the model recommended and what the team executed, and until that gap carries consequences, data-driven is just language. The model is not the problem. The organization’s relationship with accountability is.
What the conversation should actually be about
The industry has been asking the wrong question for years. The question is not which MMM to trust. It is whether your organization has shifted its expectation from what was my performance to what could my performance be.
That shift is what separates organizations using measurement as a rearview mirror from those using decision science as a forward system. A well-built closed loop, one that measures incrementally, plans forward with financial constraints, forecasts probability to goal, and reconciles execution against the plan, does not just produce better answers. It creates accountability for acting on them. It makes the gap between what the model recommended and what the team actually did visible, named, and owned.
The critics of MMM are not wrong that something is broken. They are just looking at the model when they should be looking at the system around it: the goals that drive individual behavior, the data architecture that slows the feedback loop, and the accountability structure that never required anyone to change in the first place.
Measurement is only as good as the decision it produces. The industry keeps arguing about the model. The real work is everything that comes after.
