# What is marketing mix modeling (MMM)?

> Marketing mix modeling uses weekly spend and sales to estimate what each channel adds. How it works, a simple example and the free tools from Google and Meta.
- **Author**: Marxx
- **Published**: 2026-11-20
- **Category**: Measurement
- **URL**: https://marxx.ai/library-glossary/marketing-mix-modeling

---

Marketing mix modeling (MMM) is a statistical method that uses your spend by channel and your total sales over time to estimate what each channel adds to sales. It uses totals, not data on individual people.

Google describes MMM as measuring the impact of marketing activity to guide budget decisions, using aggregated data with no cookies or user-level data ([Meridian on GitHub](https://github.com/google/meridian), checked 2026-10-20).

## What goes in

- Weekly spend per channel, such as Meta, Google, YouTube and TV
- Weekly sales
- Other things that move sales, such as festivals, price changes and discounts

The model needs spend that goes up and down over time. If you spend the same every week, it cannot tell channels apart.

## A worked example

Say a model, fitted on your weekly data, says each week's sales come from:

| Source | Weekly spend | Sales added | Return per INR 1 |
|---|---|---|---|
| Sales with no ads | none | INR 6,00,000 | not applicable |
| Meta | INR 2,00,000 | INR 4,40,000 | 2.2 |
| Google search | INR 1,00,000 | INR 2,00,000 | 2.0 |
| YouTube | INR 50,000 | INR 40,000 | 0.8 |

The numbers are made up. YouTube returns INR 0.80 per INR 1 in this model. You might move some of its budget to Meta, then check with a [geo-lift test](/glossary/geo-lift-test) before you move a lot.

## What it is good and bad at

| Good at | Weak at |
|---|---|
| The big picture across channels | Judging a single ad |
| Offline channels like TV | Fast answers; a model takes time to build |
| No tracking of people | Needing enough weeks of data |

## Free tools

Two big platforms publish open-source MMM code. Google's Meridian is an MMM framework for running your own in-house models, under the Apache-2.0 licence ([Meridian on GitHub](https://github.com/google/meridian), checked 2026-10-20). Meta's Robyn is an experimental, semi-automated, open-sourced MMM package under the MIT licence ([Robyn on GitHub](https://github.com/facebookexperimental/Robyn), checked 2026-10-20). Both need someone who can run code.

MMM answers a different question from user-level tracking. See [MMM vs MTA](/compare/mmm-vs-mta), and read [incrementality](/glossary/incrementality) for how to confirm what the model says.

## FAQ

### Do small brands need MMM?

Usually not early. With a few channels, [MER](/glossary/mer) and simple on/off tests may be enough. MMM pays off as the number of channels and the spend grow.

### How is MMM different from attribution?

Attribution follows clicks and views to credit a sale. MMM looks at totals over time and needs no click path.

### Does MMM prove cause?

Not alone. It finds patterns in your past data. Back it with a lift test.

---
- [More Measurement articles](https://marxx.ai/library-glossary/category/measurement)
- [All articles](https://marxx.ai/library-glossary)