Marketing mix modeling, explained for the CMO who has to choose a measurement model

What is marketing mix modeling?

The three questions it answers

1. Historical return on investment (ROI) and contribution from each marketing channel.
2. The response curve for each channel, meaning how impact varies with spend.
3. How to allocate future budget to get the most business outcome.

Here is an example, not a benchmark. Two years of weekly data gives 104 rows, 52 for each year. Each row carries spend for 10 channels plus the control variables, and the model returns the 3 answers above for every channel.

How long a first build takes

Data collection at 4 to 6 weeks plus review at 1 to 2 weeks puts 5 to 8 weeks of work in front of any modeling. In our reading, that data work is the part to staff and schedule first, because a model built on poor inputs produces poor outputs.

How MMM differs from multi-touch attribution and incrementality tests

The three methods get compared as if a CMO had to pick one. We think they answer different questions, and the industry bodies that publish on measurement treat them that way. The table below is our summary of how they differ.

What each one sees

Why they are complements

A worked example

Take round example inputs, not benchmarks. MTA credits paid search with 104 conversions in a month. A geographic holdout test on the same campaign finds 52 incremental conversions, so 52 of the 104 credited conversions is 50%.

The steps that follow:

1. Treat MTA's 104 as a daily optimization signal, not a causal figure.
2. Feed the test's incremental ROAS into the MMM as a prior for that period.
3. Use the calibrated MMM for the quarterly allocation you take to the CEO and the CFO.

Each method then does the job it is built for. For framing that budget conversation, see CMO, CEO and CFO alignment.

What data a marketing mix model needs, and how many weeks

Robyn, the open-source marketing mix modeling code from Meta Marketing Science, sets out its inputs in one function call. Read that call as a shopping list. It shows what finance and the media team will have to assemble before any modeling starts.

The six input groups

How many weeks

As an example, not a benchmark: 3 years at 52 weeks a year gives 156 weekly rows. A brand with 104 clean weeks has 2 years of history at 52 weeks a year. Every column in the template below needs a value in every one of those rows.

Data inventory template

What are Google Meridian and Meta Robyn?

Both are ways to build the marketing mix model (MMM) described above, and they differ in how they estimate it and what they ask of the user.

Meridian, from Google

Robyn, from Meta

Side by side

How incrementality tests calibrate a marketing mix model

Calibration is the step where a marketing mix model stops relying only on historical correlation and borrows an answer from an experiment. You run a geo or lift test on one channel, convert the measured incremental return into a prior, and the model starts from that evidence instead of a blank guess.

How the prior pulls the estimate

A worked example, with round inputs that are not benchmarks. Suppose the uncalibrated model credits paid search with $10 of revenue per $1 of spend, and a geo test measured $3 per $1: the gap is $7 per dollar. A tight prior around $3 pulls the posterior toward the test. A loose prior lets the 104 weeks of history keep more say. Where the posterior lands depends on how much uncertainty you assign the test, and that choice belongs in the model documentation your CFO reads.

Why organic demand controls move lower-funnel credit

In our reading, the control and the experiment do overlapping jobs. The control strips out baseline interest in every week of history. The experiment checks the result directly for one channel.

A calibration checklist

1. Pick the channel with the largest gap between platform-reported and model-estimated return.
2. Run a geo test with a holdout of comparable regions.
3. Convert the measured return, with its confidence interval, into a prior.
4. Add organic demand controls before reading the calibrated result.
5. Re-test a channel when its spend level or creative changes materially.

How an MMM is built and checked, step by step

Why structure choices change the answer

A worked example of the validation meeting

As an example, not a benchmark: the team fits 104 weeks of data and produces 10 candidate models. It keeps 3 that pass both fit and plausibility, so 7 of the 10 are discarded. Ask to see the 3 side by side. If they disagree on a channel's return on ad spend (ROAS), we think that channel is a candidate for an incrementality test before money moves.

A worked example of MMM results on a board slide

This example is ours, and its numbers are round illustrative inputs, not benchmarks. Picture a company that fits a marketing mix model (MMM) on 104 weeks of history, which is two years of 52 weeks. The model covers search, paid social, online video and two offline channels.

Contribution and return

Show contribution as each channel's share of the outcome the model credits to media, next to the baseline that would have happened without media. Show return on investment (ROI) as a range, never as a single point.

A channel with a wide range is one the model is unsure about. The board should hear that before it hears the point estimate.

Response curves and the reallocation

Each curve rises and then flattens. In our example, search sits past the bend, so extra spend buys little. Paid social sits on the steep part, where the model estimates the next dollar returns 3 dollars.

Move 10 thousand dollars a week from search into paid social at a marginal return of 3, and the projected gain is 30 thousand dollars a week before you subtract what search loses. Over 52 weeks, 10 thousand dollars a week is 520 thousand dollars moved. Put the net figure on the slide, after the search loss, and propose a test before committing the full year.

One-slide template

The diagnostics belong in an appendix the finance team can read. For the deck around this slide, see our board deck template. For agreeing on the numbers with your chief financial officer before the meeting, see CMO, CEO and CFO alignment.

Should you build or buy a marketing mix model?

We decide build or buy on four factors: in-house analytics headcount, data readiness, refresh cadence and budget. If two or more of them show a gap you cannot close this fiscal year, we think buying first and building later is the safer path.

A cost comparison frame

Price staff time in days. As an example, not a benchmark: if each refresh takes an analyst 10 working days and you refresh 3 times a year, that is 30 analyst days a year before any new channel or incrementality test is added. Compare that against your headcount plan and the frame in our in-house versus agency cost guide.

Ten questions to ask any provider

1. Which modeling method do you use, and can we see the code?
2. Who owns the model, the data and the outputs if we leave?
3. How many weeks of history do you require?
4. How often do you refresh, and what does each refresh cost?
5. How do you bring in experiment results?
6. How do you show uncertainty on return on investment (ROI)?
7. What do you do when a result contradicts business sense?
8. Who on your team does the modeling, and who presents to us?
9. What data do we prepare, and what do you prepare?
10. What does a handover to our own team look like?

Put these in your request for proposal and check the answers against the contract checklist.

The first 90 days

• Days 1 to 30: audit data readiness, name one owner, list every channel and its spend source.
• Days 31 to 60: send the ten questions to providers and price the open-source option in analyst days.
• Days 61 to 90: choose, sign, and agree on the first board question the model must answer.

What this page does not cover

This page does not compare or rank marketing mix modeling providers, quote prices for consultancies or software, or walk through installing and running the Meridian or Robyn code. It does not set a single minimum data requirement that fits every business. It is not financial advice, and any budget reallocation should be agreed with your finance team.

Related on CMO Tribune: In-house vs agency cost, compared with your own numbers; Marketing team structure by revenue band: functional, pods, hub-and-spoke and business-unit models; What CEOs and CFOs expect from marketing, in their own words.