The CFO's Guide to Marketing Mix Modelling

Marketing is usually one of the three largest discretionary lines on the P&L. It is also the one where the CFO has the least independent read on what the money actually bought. Attribution dashboards from the ad platforms show conversions that overlap, overstate, and change definition every quarter. Agency post-campaign decks lean on the same numbers.

Marketing Mix Modelling (MMM) is the tool built for this problem. It is a statistical decomposition of revenue into the drivers that produced it: paid media by channel, price, promotion, distribution, seasonality, and a base level. Done well, it gives finance an audit-grade answer to "what did marketing contribute" and "what happens to revenue if we spend GBP 5m less".

This guide is written for CFOs, FDs and finance business partners who sign off marketing budgets. It covers what MMM actually measures, how to tell a credible programme from an expensive dashboard, and the questions to ask before the next budget round.

Why attribution is not enough for finance

Every ad platform reports its own conversions. Google, Meta, TikTok, Amazon and the affiliate network will each cheerfully claim the same online sale. Totalled up, platform-reported conversions routinely exceed real revenue by two to five times.

The mechanism is straightforward. Each platform runs its own last-touch or data-driven attribution model on the events it saw. None of them sees the others. None of them models the base demand that would have existed with no advertising. And the definitions (view-through window, deduplication, modelled conversions) shift silently as the platforms update their products.

For a marketing team that is optimising within a channel, platform attribution is useful. For a CFO deciding whether marketing is worth GBP 20m next year, it is not. The numbers do not tie to the P&L and cannot be audited.

What MMM actually does

MMM approaches the question from the top down. It takes weekly (or daily) revenue and regresses it against everything that could have caused it to move: media spend by channel, price, promotion depth, distribution changes, competitor activity, seasonality, weather, and macro factors. The output is a set of coefficients that decompose revenue into contributions from each driver, with a base level for demand that would have existed anyway.

Because the total is anchored to actual revenue, the contributions add up to 100% of the number finance already signs off. There is no double counting across channels. Every channel's contribution is expressed with a confidence interval, so a CFO can see not just the point estimate but how much the model actually knows.

The important consequence: MMM answers questions attribution cannot. What is the return on the brand campaign that generates no clicks? What is the incremental value of price promotion on top of media? What would happen to revenue if we cut TV by 30%? Attribution has no view on any of these. MMM does.

The three numbers finance should ask for

A useful MMM programme delivers three outputs. If a vendor cannot produce all three, the programme is not fit for finance sign-off.

1. Base vs incremental revenue. What proportion of revenue would have happened with no marketing at all? For most established brands this is 60-85%. It is the single most important number in the model and the one most vendors bury.
2. Channel ROI with confidence intervals. Every channel's return per GBP spent, with a range. A point estimate with no range is not a statistical answer.
3. Response curves and saturation points. Where does each channel stop paying back? This is what turns MMM from a report into a decision tool: it tells you where the next GBP of budget will earn most.

What "modern" MMM looks like

The MMM of the 2010s was an annual exercise: a six-month build, a slide deck, a filing cabinet. Modern MMM is different in three ways that matter to finance.

Bayesian priors constrained by experiments. The model is anchored to what is already known from geo holdouts and platform lift tests, which kills the implausible TV ROI of 47 results that used to embarrass the practice.

Monthly refresh. Code-first pipelines mean the model updates alongside the management accounts rather than once a year. Reallocation decisions can be taken in the actual planning cycle.

Narrower scope. Good modern MMM resists the temptation to model everything. It scopes to the decisions finance and the CMO actually need to take, and accepts that the rest is noise.

How to govern the programme

Treat MMM like any other investment-grade measurement. The controls are not exotic; they are the ones a CFO would apply to a valuation model or a demand forecast.

• Written methodology. How was every variable defined, transformed and validated. Filed, versioned, auditable.
• Out-of-sample validation. The model is fitted on all but the most recent 8-12 weeks, then tested on the held-back window. Prediction error is published, not hidden.
• Confidence intervals on every ROI. Point estimates without ranges are the single most common tell of a vendor selling a dashboard rather than a model.
• Experimental validation. At least one geo holdout or platform lift test per year, on the largest channel. The MMM's read on that channel is compared to the experiment. If they disagree materially, the model is retuned.
• Independent reporting line. The team that runs the model should not report into the media team whose spend it is judging. Sit it in finance, in an effectiveness function, or with an independent partner.

The commercial case

For a GBP 20m annual media budget, a well-run MMM programme typically costs GBP 150-400k a year. The value comes from two sources.

The first is avoided misallocation. Most media plans have 10-20% of spend on channels sitting on the flat part of their response curve: money that generates almost no incremental revenue. Moving that budget to channels still on the steep part is worth 10-20% of media spend in incremental revenue, all else equal. On GBP 20m that is GBP 2-4m per year.

The second is defensibility. When the board asks whether marketing should be cut, MMM lets finance answer with a range and a set of assumptions rather than a marketing team's assertion. In three of every four downturns we see, MMM protects a marketing budget that would otherwise have been cut too far, or right-sizes one that had grown past the point of return.

What to look for in a vendor

The single most useful screen is to ask a vendor to walk through a model they have built. Not a case study, not a dashboard: an actual model, with the assumptions, the data, and the out-of-sample fit. Teams that pivot to demoing a UI are the ones to avoid. Teams that spend forty minutes on data engineering and variable selection are the ones to shortlist.

Ask specifically:

• What is your base rate and how did you validate it
• What is your out-of-sample MAPE on the last three refreshes
• Which experiments have you used to constrain the priors
• Show me a channel where your model changed a client's mind, and what happened next

The answers will tell a CFO in twenty minutes whether the programme is investment-grade or a dashboard with regression underneath.

The bottom line

Marketing is auditable. It has been auditable for decades. The reason so many finance teams still take marketing performance on trust is that the industry allowed platform attribution to fill the gap, and attribution is not a finance tool.

MMM, done to the standard finance would apply to any other model, closes that gap. It is not exotic, not new, and not expensive relative to the budget it governs. It is how a CFO gets an independent answer to the question the board keeps asking: is the marketing budget the right size, and is it in the right places.