Media mix modeling (MMM) is a statistical method that estimates how much each marketing channel contributes to sales by regressing aggregate outcomes (weekly revenue, for example) on spend or exposure per channel, while controlling for seasonality, price and other factors. It needs no user-level tracking, which makes it privacy-durable and able to measure offline media.
What is media mix modeling?
Advertising rarely works instantly or linearly: its effect lingers after the spend and each extra dollar buys less. Google researchers proposed a Bayesian MMM with flexible functional forms for exactly these carryover and shape effects, and showed how to compute ROAS and marginal ROAS from it.2Source 2 · Google Research, 2017Bayesian Methods for Media Mix Modeling with Carryover and Shape Effectsresearch.google Google’s open-source Meridian models lagged effects that taper off over time with an adstock function and diminishing marginal returns with a two-parameter Hill function.1Source 1 · Google Meridian documentationMedia saturation and laggingdevelopers.google.com
Sales(t)=Baseline(t) + Σ channels β × Saturation(Adstock(Spend(t))) + Controls(t) + ε
- Baseline
- Sales you would get with no media: trend and seasonality
- Adstock
- Weighted carryover of past spend into this period
- Saturation
- Diminishing-returns curve, for example a Hill function
- Controls
- Price, promotions, distribution, macro factors
Worked example: reading an MMM output
| Channel | Quarterly spend | Modeled contribution | Modeled ROAS | Marginal ROAS |
|---|---|---|---|---|
| Paid search | $300,000 | $900,000 | 3.0 | 1.4 |
| Paid social | $250,000 | $550,000 | 2.2 | 1.9 |
| Online video | $150,000 | $270,000 | 1.8 | 1.7 |
| Baseline (no media) | — | $2,600,000 | — | — |
The marginal column is the actionable one: budget should flow toward the channel where the next dollar earns most, not the one with the best average.
Common mistakes
- Too little history or variation. If a channel’s spend never changed, the model cannot learn its effect.
- Uninformative priors on small data. Google’s paper found that with small samples the priors strongly shape the results.2Source 2 · Google Research, 2017Bayesian Methods for Media Mix Modeling with Carryover and Shape Effectsresearch.google
- Never validating. Calibrate MMM coefficients against incrementality tests and check holdout accuracy.
- Treating outputs as precise. Report intervals, not single numbers.
How to prepare MMM data in Kimo
The hard part of MMM is the dataset: weekly spend, impressions and outcomes, consistently defined across every channel. Kimo builds that table from Google Ads, Meta, TikTok, YouTube and your revenue source, exports it to BigQuery or Snowflake for Meridian or Robyn3Source 3 · Meta Marketing Science (GitHub)Robyn: semi-automated marketing mix modelinggithub.com, and brings modeled contributions back into the Command center. Our unified marketing measurement whitepaper covers how MMM, attribution and lift tests fit together.
Frequently asked questions
What is the difference between MMM and multi-touch attribution?
How often should an MMM be refreshed?
Is MMM only for large advertisers?
Sources
3 references- Media saturation and lagging (opens in a new tab)Google Meridian documentationdevelopers.google.com
Adstock for lagged effects; Hill function for saturation.
- Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects (opens in a new tab)Google Research2017research.google
Carryover and shape effects; ROAS and mROAS; priors dominate in small samples.
- Robyn: semi-automated marketing mix modeling (opens in a new tab)Meta Marketing Science (GitHub)github.com
Open-source, semi-automated MMM package from Meta Marketing Science (R and Python): ridge regression, evolutionary hyperparameter search, budget allocation.
External sources were accessed at the time of writing. Kimo product details, customers and figures in examples are illustrative unless a source is cited.
