What media mix modeling does
MMM looks at history from above. It takes two or three years of spend per channel, total sales, and outside factors such as seasonality, promotions and prices, and fits a model that explains sales from those inputs. The model then says how much revenue each channel added, and what would happen if you moved budget between them.
Because it works with totals, not people, MMM is unaffected by cookie limits, ad blockers or lost click IDs. It can measure channels that leave no click to track: TV, radio, podcasts, billboards, sponsorships. That is why it came back into fashion as user-level tracking got harder.
Marketing mix modeling is the older and broader name. It can also model price, distribution and product changes. Media mix modeling focuses on the media budget. In practice, people use the two names and the abbreviation MMM interchangeably.
How MMM works
At its core is a regression: sales in each week are explained as a baseline plus a contribution from each channel.
Sales(t) = baseline + Σ βᵢ × saturation(adstock(spendᵢ)) + seasonality + other factors + error
- Adstock models the delay: an ad seen this week still nudges sales next week, fading over time. A podcast sponsorship may keep working for months; a search ad mostly works the same day.
- Saturation models diminishing returns: the first $10,000 a week on a channel buys more than the next $10,000. This is what lets the model recommend where the next dollar should go.
- Baseline is the revenue you would make with no paid media: brand, word of mouth, organic search, returning customers.
Modern MMMs are often Bayesian. You give the model prior beliefs, for example results from an incrementality test, and it combines them with the data. Two open-source options are widely used: Google's Meridian and Meta's Robyn.
Worked example
A direct-to-consumer brand fits an MMM on two years of weekly data. It averages $100,000 of revenue a week. The model's decomposition for the last quarter:
| Source | Share of revenue | Weekly revenue | Weekly spend | Modeled ROAS |
|---|---|---|---|---|
| Baseline | 55% | $55,000 | — | — |
| Meta Ads | 20% | $20,000 | $8,000 | 2.5 |
| Google Search | 15% | $15,000 | $4,000 | 3.75 |
| Podcasts | 10% | $10,000 | $5,000 | 2.0 |
Two things stand out. First, more than half of revenue doesn't depend on paid media. Last-click attribution would have handed most of that to whatever ad was clicked last. Second, the saturation curves show Meta near its flat part: the model predicts an extra $2,000 a week there adds about $2,600 of revenue, against about $5,500 if it went to search. That is the budget question MMM is built to answer.
MMM vs multi-touch attribution vs incrementality tests
| MMM | Multi-touch attribution | Incrementality test | |
|---|---|---|---|
| Data | Aggregate spend and sales over time | User-level journeys | A controlled experiment |
| Measures | Estimated contribution per channel | Credit split across touches | Causal lift of one change |
| Offline channels | Yes | No | Yes, with geo tests |
| Speed | Months of history, refreshed weekly or monthly | Daily | Weeks per test |
| Weak spot | Correlation, needs lots of history and variation | Only sees tracked touches, not cause | One answer per test |
They work best together. Use multi-touch attribution or first-touch attribution for day-to-day campaign decisions, incrementality tests to check the biggest channels, and MMM to set the yearly budget split, calibrated with the test results.
What you need to build an MMM
- Two to three years of weekly data, or at least a year of daily data. Less history makes the estimates unstable.
- Spend (or impressions) per channel for every period, from the ad platforms and invoices.
- The outcome from your own records: revenue or new customers from your payment provider, not platform-reported conversions.
- Variation. If you spent the same on every channel every week, the model can't tell them apart. Planned changes in spend make better models.
- Control variables: seasonality, holidays, price changes, launches, promotions.
Common MMM mistakes
- Too little data for too many channels. Ten channels and one year of weekly data is 52 rows for a dozen unknowns. Group small channels or wait.
- Channels that always move together. If Meta and Google budgets rise and fall in step, the model can't separate them.
- Leaving out what drove sales. A big launch missing from the inputs gets credited to whatever channel spent most that month.
- Treating the output as fact. An MMM is an estimate. Check its biggest claims with an incrementality test before moving large budgets.
How MetricRaven fits in
MetricRaven doesn't build media mix models. It credits each payment to the visit that first brought the customer, which is attribution, not MMM. What it can give an MMM is clean inputs from your own data:
- Revenue per day, week or month from your payment provider, net of refunds, split into new and renewal, from the Reports API (
GET /stats/revenueandGET /stats/traffic). - Daily spend, clicks and first-touch revenue per campaign for connected Meta Ads and Google Ads accounts (
GET /stats/adswith a campaign). - Chart notes for launches and promotions, so you know which weeks need a control variable.
See the Reports API and Revenue dashboard docs.
Questions
- What is the difference between MMM and multi-touch attribution?
- MMM uses aggregate spend and sales over time to estimate each channel's contribution, without tracking people. Multi-touch attribution follows individual journeys and splits credit between tracked touches. MMM covers offline channels; attribution is faster and more detailed for digital ones.
- How much data do you need for media mix modeling?
- Usually two to three years of weekly data, or at least a year of daily data, with real variation in spend per channel. With less, the estimates swing too much to act on.
- Are there free media mix modeling tools?
- Yes. Google's Meridian and Meta's Robyn are open-source MMM packages. They still need clean spend and revenue data and someone comfortable with statistics to run them.
- Is marketing mix modeling the same as media mix modeling?
- Mostly. Marketing mix modeling is the broader term and can include price, distribution and product. Media mix modeling focuses on the media budget. Both are called MMM.