At CIMM Summit XV, I joined a panel on “The Future of MMM and Agentic AI,” moderated by Angelina Eng, former head of measurement at IAB, with leaders from across the martech space. We disagreed on plenty, but the he one thing nobody pushed back on was that most measurement still ends up as a report when it should be driving a decision.
Marketing mix modeling exists to answer a decision: what should we do next. In practice, most teams get an autopsy instead. The model runs once a quarter, confirms what finance already suspected about last period’s spend, and gets filed away until the next readout. That’s not a decision tool. That’s a report card.
The gap matters most for the spend that’s hardest to defend: brand and upper-funnel investment. Performance channels get judged on this week’s conversions, so they survive budget reviews almost by default. Brand media doesn’t have that luxury. Its payoff shows up over quarters and years, which means it needs a measurement approach built to see causal, long-horizon effects, not just last month’s attribution-friendly clicks. If your MMM only looks backward, brand is the first line item to get cut when someone asks “prove it worked.”
Why lookback models persist
Three habits keep MMM stuck in report-card mode:
- Slow refresh cycles. Quarterly or annual model runs mean the output is stale before anyone acts on it.
- Correlation dressed up as causation. A model fit purely to historical data will happily explain the past while being wrong about what happens if you change the plan. That’s how marketing mix modeling earned a reputation as a deck exercise instead of a planning input.
- No link from measurement to action. Even when a model is directionally right, if it doesn’t feed a weekly plan, it’s just an interesting slide.
What a causal, forward-looking model actually changes
The fix isn’t a better chart. It’s a different question. Instead of “what happened,” the model needs to answer “what happens if we change this.” That’s the difference between forecast accuracy and causal MMM: a model can be great at explaining history and still useless for predicting the result of a decision you haven’t made yet.
Getting to causal, decision-grade output usually means two shifts:
- Calibrated priors instead of cold-start regression. A model that starts with test-calibrated priors from real experiments and outcomes-based activations, then updates on your own data, is far more defensible than one asked to find causality from historical variance alone.
- A cadence that matches how budgets actually move. Quarterly readouts can’t keep up with weekly buying decisions. If the model doesn’t refresh on the same clock as the plan, it will always be describing a world that’s already changed.
Giving brand its due
Upper-funnel measurement specifically needs a model willing to represent multi-year brand effects, not just the 4-6 week window most attribution-adjacent tools default to. That means explicit response curves for brand channels, diminishing returns modeled over the right time horizon, and a way to show the probability of hitting a goal under a given plan, not just a single point estimate. Our breakdown on how to measure brand awareness gets into the mechanics, but the principle is simple: if your model can’t separate short-term activation from long-term brand building, it will systematically undercount brand’s contribution, and the budget will follow that error.
This is the problem we built Keen AI Cortex to solve. Its calibrated-priors component, the Marketing Elasticity Engine, grounds the model in outcomes-based evidence rather than starting from a blank slate, so it can represent causal brand effects without requiring an always-on testing program to get there.
The real test
Ask your current MMM one question: if I move $500K from search to streaming next month, what happens to revenue, and how confident are we? A report card can’t answer that. A decision system can. If your model only tells you what already happened, you’re not measuring marketing, you’re grading it.
