Marketing measurement

Why CFOs don’t trust marketing mix models — and what to do instead

Marketing mix modeling (MMM) promises to answer one of the most expensive questions in any company: how much does each marketing channel really contribute, and where should the next euro go? Yet in most boardrooms the answer is received with polite scepticism. A BCG study found that while around 75% of companies have marketing mix models to measure and optimize marketing, only 26% actually use them to take budget allocation decisions.

That gap is not a communication problem. It is a design problem.

What a marketing mix model is supposed to do

An MMM is a statistical model that links business outcomes — sales, sign-ups, revenue — to marketing spend by channel, alongside other factors such as seasonality, price or macro conditions. Once fitted, it estimates the incremental impact of each channel and feeds a budget optimizer that recommends where to move money.

The core problem: correlation, not causation

The hard truth is that media typically drives only 10–20% of business outcomes, and a single channel generally somewhere in the 0–10% range — usually at the lower end. Everything else comes from demand, product, pricing, distribution, competition and the economy.

Because media explains such a small share of the outcome, many very different models fit the data almost equally well. Shift a little of the explained variation from external factors to media and the calculated impact of a channel can change dramatically. When fit or prediction accuracy improves, it is most often because external factors are better captured — not because media is better measured.

A better fit does not mean you have found the real impact of a channel.

Why the numbers get “tuned”

Providers know this. Model structure (in frequentist MMMs) or priors (in Bayesian ones) can be adjusted until the result is accepted by stakeholders — with no certainty that it represents reality. CFOs are very aware of it: many have seen two completely different MMMs produced for the same business at the same time. The rational response is to discount them entirely.

The same pattern appears outside marketing. Sales business plans and product-launch assessments are usually department-owned analyses that miss the wider context. Each one tends to overstate its own impact — not out of bad faith, but because a model that ignores collinear factors (a product launch running at the same time as a campaign, for instance) attributes their effect to the drivers it does include. As one CFO famously put it: “If I added up all the impacts assumed by each business function owner, we’d end up with 3x higher revenue.”

Stop choosing “the” model

Correct identification is the central problem, and it cannot be solved reliably with the standard MMM toolkit. So instead of pretending to find the one right model, we explore the whole range of plausible explanations and look for families of models that tell a similar story about how the business works — and point in the same direction.

In practice we test between 700 and 2,000 model variations for each business outcome, varying the assumptions that matter most:

• How media interacts with underlying demand
• How click and impression price changes are treated
• How saturation is handled — informed by attribution data or discovered inside the model
• The assumed scale of total media impact (small, medium or large)
• How much weight recent data receives
• Different adstock (carry-over) functions
• Which channels, sub-channels and external factors are included
• Whether brand is allowed to have an indirect impact

Around 90% of the compute goes to testing, validation and optimization rather than to fitting a single model. The focus is on the validity of the recommendation, not of any one model.

From models to decisions, the way a CFO decides

If 100 models all agree that you should reduce TV by 700k–1.5M and increase Social by 400k–700k, that is far more useful to a CFO than a single model saying “reduce TV by exactly 700k and increase Social by 700k”. When the competing explanations agree, you can move with confidence. When they disagree, you know exactly where to run a lift or incrementality test — or you can take an informed, explicit bet.

We then risk-adjust the recommendation: large moves that look ROI-positive on average but carry a lot of uncertainty are penalized, just as a CFO would treat any other investment. The output is a range, not a false point estimate — “increase by 20–40%, the lower end if you believe X, the higher end if you believe Y”.

Speaking the CFO’s language

Recommendations are expressed in the universal language of finance — risk-adjusted returns and profit — rather than in department-specific jargon. And because the model covers the full commercial engine, its recommendations can be back-tested with a gap analysis against what actually happened.

This is the idea behind a commercial digital twin: one consistent model of the business in which marketing, sales, product and planning questions all get answered from the same foundation.