Marketing mix modeling vs. incrementality testing
By Harry Burman · Updated October 15, 2026
Marketing mix modeling (MMM) uses your sales and spend history to estimate how much each channel contributes. Incrementality testing changes one channel on purpose and measures what happens. Models cover everything at once but can confuse coincidence with cause. Tests prove cause but cover one thing at a time. They work best together.
Marketing mix modeling, also called media mix modeling or MMM, is a statistical model of weekly sales against weekly spend by channel, plus price, promotions and season. It estimates every channel at once from history, but it can confuse coincidence with cause, for example when a brand spends more before a holiday that would have sold well anyway. Incrementality testing changes one channel on purpose, usually by pausing ads for a holdout group of people or regions, and measures what that channel caused. A test proves cause but covers one channel at one spend level at a time. Brands spending across many channels with years of data can plan with MMM and calibrate it with tests on their biggest channels. Younger brands with one or two channels should start with a test.
What is marketing mix modeling?
Marketing mix modeling, also called media mix modeling, is a statistical model of your sales over time. It takes weekly sales and weekly spend by channel (Meta, Google, TV, Amazon ads and so on), plus things like price, promotions and season, and estimates how much of the sales each one explains.
Modern MMMs build in two facts about advertising (Jin and others, 2017):
- Carryover. An ad's effect lasts beyond the week it runs, and fades over time. This is often called adstock.
- Diminishing returns. Each extra dollar in a channel does less than the one before. The marginal ROAS calculator shows the idea on your own numbers.
sales each week ≈ base sales
+ effect of channel 1 spend (with carryover and diminishing returns)
+ effect of channel 2 spend …
+ price, promotions, seasonOpen-source tools have made MMM much more accessible. Meta publishes Robyn, and Google publishes Meridian.
What is incrementality testing?
An incrementality test is an experiment. You change one channel for one group, usually by pausing ads for a holdout group of people or regions, and compare sales with a group where nothing changed. The difference is what that channel caused during the test. Geo tests and platform conversion lift studies are the common forms. See what is incrementality for the basics.
How they compare
| Marketing mix modeling | Incrementality test | |
|---|---|---|
| Evidence | Patterns in history | A change made on purpose, so it measures cause directly |
| Coverage | Every channel at once | One channel, at one spend level, at a time |
| Data needed | A long history of weekly spend and sales across channels, with enough variation in spend to learn from | A few months of sales history and a few weeks of running |
| Cost | Analysis time | The sales lost while ads are paused, if the ads work |
| Main risk | Crediting a channel for something that happened at the same time, like a seasonal rise | Too small to see a realistic effect |
Why models need tests
MMM has to infer cause from history, and history is tangled. Brands often spend more when sales are expected to be high, like before a holiday. A model then sees spend and sales rise together and can't easily tell how much the ads did and how much the season did.
Experiments untangle this. A study of large ad experiments at Facebook compared their results with non-experimental methods applied to the same campaigns, and found those methods often missed the true effect, sometimes by a lot (Gordon and others, 2019).
That's why both major open-source tools are built to take experiment results:
- Robyn can be calibrated with lift results, pulling the model's estimates toward what tests found.
- Meridian is Bayesian and can use experiment results to set its starting beliefs (priors) about each channel.
Which should you use?
- Spending across many channels, with years of data? MMM gives a full picture for planning. Calibrate it with tests on your biggest channels.
- One or two channels, or a young brand? There usually isn't enough history for MMM. Start with a test on your biggest channel.
- Need to know if one channel pays off? Test it. Then compare the incremental ROAS with your break-even ROAS.
- Sales split across your own site and Amazon? Make sure whatever you use counts both. Platform tests often miss Amazon. Geo tests and MMMs built on total sales don't have to.
Where Spillmark fits
Spillmark runs incrementality tests: geo holdouts that measure how many Amazon sales Meta ads cause. Each test is recorded with its design, result and uncertainty, and later tests and recommendations build on your brand's own past results. Those results can also calibrate an MMM if you run one. The method is on how lift is calculated.
Frequently asked questions
What is marketing mix modeling?
A statistical model of sales over time that estimates how much each marketing channel, plus price, promotions and season, contributes to sales.
Is MMM better than incrementality testing?
They do different jobs. MMM covers every channel from history; tests prove what one channel causes. The strongest setups use tests to calibrate the model.
What is MMM calibration?
Using experiment results to adjust a model's estimates. Robyn can calibrate with lift results, and Meridian can use them to set its priors.
How much data does marketing mix modeling need?
A long history of weekly sales and spend by channel, with enough ups and downs in spend for the model to learn from. Young brands often don't have enough.
Can MMM measure Meta's effect on Amazon sales?
Only if the model's sales include Amazon. A model built on site sales alone will miss it, just as platform reports do.
Sources
- Jin, Wang, Sun, Chan & Koehler, Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects (Google, 2017)
- Robyn, Meta's open-source marketing mix modeling package
- Meridian, Google's open-source marketing mix modeling framework
- Gordon, Zettelmeyer, Bhargava & Chapsky, A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook (Marketing Science, 2019)