
Marketing mix models (MMM) estimate each channel’s contribution to sales from aggregate data, without relying on individual tracking. Open-source tools, such as Meridian (Google) and Robyn (Meta), have made this approach accessible to small teams, provided there is quality data, realistic expectations and technical support.
The essentials in a definition
Open-source marketing mix is the application of open statistical models, such as Meridian and Robyn, to aggregate data to estimate each channel’s contribution. It requires quality data, technical support and validation with experiments, and can use platforms such as BigQuery to organise data.
What it is and why it is back
An MMM relates weekly or daily sales to investment by channel, prices, promotions and external factors, to estimate the effect of each and the response to more or less investment. It has regained relevance because it works with aggregate data, respecting privacy, and gives a cross-channel view. Meridian and Robyn are open-source projects with public documentation.
It does not replace incrementality tests; it complements them.
Data needed
Consistent time series, generally two or more years: sales (or leads), investment by channel, prices, promotions, seasonality, events and relevant external variables. Quality matters more than quantity: reliable data, at the same granularity, with no gaps. A data warehouse, such as BigQuery, can help organise them.
If data are scarce, the model will be uncertain; accept it.

An open MMM, simple and validated
- Gather clean, consistent data
- Configure the model with technical support
- Validate with tests and business knowledge
- Decide in gradual steps and repeat
How to use realistically
Results are estimates with uncertainty intervals, not truths. Compare models with business knowledge and with experiments (holdout tests), and use them to decide gradual budget changes, not jumps. An experienced analyst is useful to configure, validate and interpret.
Record hypotheses, parameters and model versions.

A useful model says where to investigate and where to test, it does not replace judgement.
Integrate into decisions
Repeat the model periodically (for example, quarterly), track the stability of estimates and document decisions made. Communicate uncertainty to management. If a channel seems valuable in the model but is not confirmed in tests, give more weight to the tests.
Use the model to ask “what if”, not to get the final answer.
Decision table
| Element | Good practice | Risk |
|---|---|---|
| Data | Long, consistent series | Gaps and definition changes |
| Model | Configured and documented | Black box |
| Validation | Compare with holdout tests | Relying only on the model |
| Decision | Gradual changes | Budget jumps |
Common mistakes to avoid
- Using little data. Estimates become unstable.
- Treating the result as truth. There is uncertainty and assumptions.
- Not repeating. The market changes.
How to measure the result
Track stability of estimates, agreement with tests and the effect of budget changes.
Limits and exceptions
Models depend on assumptions and data, and can mislead with little data or highly correlated channels. They require skills and time. Results are estimates and should be read with caution.
Illustrative example
Illustrative example: an e-commerce SME gathers two years of weekly sales and investment data, configures an open model with an analyst’s help and obtains estimates by channel with intervals. It validates one channel’s with a holdout test and adjusts budget in small steps, reviewing the model every quarter.
In practice: checkpoints
Before building a model, write the business questions it must answer and the decisions it will inform. Check data quality first, test the model on a past period where you know what happened and keep a log of versions and parameters. Present results with ranges, and confirm the main conclusions with a small experiment before acting on them at scale.
Frequently asked questions
Do I need a data team?
It helps to have someone with statistical skills; tools simplify but do not remove the need for analysis.
How much data is needed?
Generally two or more years of consistent weekly or daily data.
Does MMM replace tests?
No; it complements them.
Which tool should I choose?
Compare documentation, community and fit to your data.
How often should I update?
Quarterly, or after relevant changes in the mix.
Use open MMMs with clean, consistent data, technical support, validation by experiments and gradual budget changes, treating results as estimates.
Sources and recommended reading
Last editorial review: October 2026. Informational content; always confirm current conditions in the official sources before deciding.