What is a media mix model?
A media mix model (MMM) is a statistical model that estimates how much each advertising channel adds to your sales, on top of what you would have sold anyway. It works from totals over time, like weekly spend by channel next to weekly sales, so it doesn't need to track individual people. That makes it one of the few ways to measure channels nobody clicks, like TV, and sales that happen in a store.
Key takeaways
- An MMM splits your sales into a baseline (what you'd sell with no ads) and what each channel added on top.
- It uses totals over time, not user-level tracking, so it can see in-store sales and channels like TV, radio and out of home.
- It accounts for carryover (ads keep working after they run) and diminishing returns (each extra dollar adds a little less).
- Its best use is planning: estimating where the next dollar will earn the most.
- It's a model estimate, not an experiment. Read its outputs as ranges, and check them against tests when you can.
What a media mix model measures
Every business sells something without advertising. Repeat customers come back, people walk by, word gets around, and some months are simply busier than others. A media mix model calls that the baseline. Its job is to estimate what advertising added on top of the baseline, and how much of that each channel is responsible for.
To do that, it looks at how sales moved over time alongside spend (or impressions) in each channel. It also looks at the things that move sales on their own: seasons, holidays, prices and promotions, and the wider economy. When spend in a channel rises and sales rise with it in a way those other factors can't explain, the model gives that channel credit.
The answer comes back in the units you care about: sales or revenue, not clicks. That's what makes an MMM useful to a finance team as well as a marketing team.
How it works, in three steps
- Line up the data. Weekly (or daily) sales next to spend by channel, plus the outside factors that move sales without any ads.
- Separate what the ads added. The model estimates the baseline and each channel's share of the rest. It also learns carryover: an ad seen this week can still bring in a sale weeks later, so a channel's effect is spread over time.
- Move the next dollar. Each channel gets a response curve. The curves flatten as spend grows, because the tenth dollar in a channel usually adds less than the first. Comparing the curves shows where the next dollar should earn the most.
The illustration below walks through the same three steps. The shapes are made up to show the idea. They are not any client's data.
-
Line up the data
The model looks at weekly sales next to weekly ad spend in each channel. It also takes in the things that move sales on their own, like seasons, holidays and promotions.
-
Separate what ads added
It estimates how much of sales would have happened anyway (the baseline) and how much each channel added on top. Ads keep working for a few weeks after they run, and the model allows for that.
-
Move the next dollar
Past a point, each extra dollar in a channel earns less than the one before. The model shows where the next dollar earns more, and that guides where the budget goes.
The split between baseline and ads is the model's estimate from the data, not a controlled test.
What a media mix model needs from you
- All of your sales. Online and in store, by week, ideally split by region. If the model only sees online sales, it can only measure online sales.
- Spend and impressions by channel. By week, for every paid channel, including offline ones like TV, radio, print and out of home.
- A calendar of what else happened. Promotions, price changes, store openings and closings, product launches, and anything unusual.
- Some variation. If spend in a channel never changes, the model can't tell what that channel does. Flights, pauses and budget shifts all help.
What it can't do
It isn't an experiment. An MMM reads patterns in history, with controls for the other things that move sales. That is an estimate, not proof, and every estimate has a range around it. Bayesian models like Google's Meridian report that range directly, which keeps everyone honest about how sure the model is.
It needs history. A brand-new channel, or a partner that only runs a few weeks a year, may not have enough data for the model to read on its own. Channels that always rise and fall together are also hard to tell apart.
It sees the short and medium term best. Effects that take many months to show up, like slow brand building, are hard for any model to pin down.
It works at the channel level. It won't tell you which keyword or which ad worked best. Platform reports and A/B tests are better for that.
MMM vs. attribution
Attribution follows individual people from ad to purchase and splits the credit between the ads they touched. It's detailed and fast, but it only sees what can be tracked, and it can't tell which of those sales would have happened anyway. An MMM sees everything, including offline sales and channels nobody clicks, but only in totals. Most teams need both. We explain the difference in incrementality vs. attribution.
Our media mix models
We build media mix models with Google's open-source Meridian. For a Florida retailer that makes most of its sales in stores, we built one that measures advertising against all sales, in store and online, and controls for the macro environment. We update it every month, and the client plans the mix with it. The model informs the calls. The people who buy each channel still make them.
The model estimates the mix as a whole adds more than $10M a year in incremental revenue. It was the first time the client could see that their advertising was working. We used it to make the case for spending more, and the next year's digital ad plan came in more than 20% higher. See the retail case.
Common questions
Is a media mix model the same as a marketing mix model?
Mostly, yes. People use both names for the same method, and both are shortened to MMM. "Marketing mix" sometimes also covers price, promotions and distribution, not just media.
How much data does a media mix model need?
Enough history for the model to see spend in each channel go up and down across different seasons. As a rule of thumb, two or more years of weekly sales and spend is a common starting point. With less history, or fewer changes in spend, the estimates get less certain.
How often should a media mix model be updated?
Often enough to keep up with how you plan. The retail model on our Results page is refreshed every month, so the read on each channel stays current for the next planning call.
Can a media mix model prove that advertising works?
No model can prove it on its own. An MMM gives a careful estimate, with a range around it. Controlled tests, like a lift study or holdout, are the stronger proof for one channel at a time.
