Quick answer
Media mix modeling (MMM) is a statistical method that estimates how much each marketing channel contributed to sales by regressing weekly sales on channel spend, seasonality and other controls. It needs no user-level tracking, but it needs about two years of weekly data with real swings in spend.
Formula: Sales(t) = baseline + Σ channel effect_i(adstock, saturation of spend_i) + controls + error
What is media mix modeling?
In marketing measurement, media mix modeling (MMM) is a statistical method that estimates how much each marketing channel contributed to sales by regressing a business outcome, such as weekly revenue, on the spend or exposure of every channel over time, plus controls like seasonality, price and promotions. It works on aggregated time series, so it needs no cookies, user IDs or click paths.
That is its appeal and its cost. Because it never looks at individual users, MMM can include TV, radio, out-of-home, podcasts and other channels that leave no click. But it can only learn from variation in the data: if you spent roughly the same on every channel every week, there is nothing for the regression to separate. The thesis of this entry is simple: MMM needs long weekly history with real swings in spend; when you do not have that, a spend table plus several attribution models side by side is the more practical first step.
The terms "media mix model" and "marketing mix model" are used almost interchangeably. Strictly, a marketing mix model also covers price, distribution and product changes, while a media mix model focuses on paid media. Both answer the same question: what would sales have been without each channel?
The basic model: baseline, channel effects, adstock and saturation
At its simplest, an MMM says:
Sales in week t = baseline + Σ (effect of channel i in week t) + controls + error
- Baseline is what you would sell with zero media: brand strength, repeat customers, organic demand.
- Channel effects are the extra sales driven by each channel's spend or impressions.
- Controls absorb things that are not media, such as holidays, price changes, competitor launches or stock-outs.
Two transformations make the channel effects realistic:
- Adstock (carryover). Advertising keeps working after the week it ran. Adstock adds a decaying share of past weeks' spend to this week's, for example 50% of last week's plus 25% of the week before.
- Saturation (diminishing returns). Each extra dollar in a channel buys less response than the previous one. A curve, often a Hill or exponential function, flattens the effect as spend rises.
Modern open-source tools, such as Meta's Robyn and Google's Meridian, fit these models with regularized or Bayesian regression, so that prior knowledge and uncertainty ranges are part of the output rather than a single point estimate.
MMM vs multi-touch attribution
| Media mix modeling | Multi-touch attribution | |
|---|---|---|
| Unit of data | Weekly (or daily) totals per channel | Individual conversion paths |
| Needs user tracking | No | Yes: sessions, clicks or user IDs |
| Offline channels (TV, radio, print) | Included | Usually invisible |
| Answers | Incremental sales per channel, response curves, budget scenarios | How credit is shared among the touches people actually had |
| Data needed before it works | About two years of weekly history | A few weeks of tracked conversions |
| Speed of feedback | Monthly or quarterly refresh | Daily |
| Main weakness | Needs spend variation; confounded by seasonality and promotions | Credits correlation, not causation; misses untracked touches |
The two are complements, not rivals. Attribution tells you who touched the path; MMM tells you what the channel added on top of the baseline. When the two disagree sharply, an incrementality test (a geo holdout or a paused campaign) is the tiebreaker. For the attribution side in depth, see multi-touch attribution and the position-based model.
What data does MMM need?
- History. Robyn's analyst guide asks for a minimum of two years of weekly data, so the model sees each season at least twice.
- Granularity. Weekly is the recommended level; daily is acceptable but noisier and needs extra validation.
- Observations per variable. The same guide recommends roughly 10 observations for every independent variable. A model with 8 channels and 4 controls needs about 120 weeks.
- Spend variation. Each channel needs weeks of higher and lower spend. Flighting (on and off weeks) or a deliberate test period helps the model a great deal.
- Controls. Holidays, price changes, promotions and distribution changes, each as its own weekly column.
- One outcome. Revenue, orders or sign-ups, defined the same way across the whole history.
Worked example: 8 weeks of spend and one regression
Sample data: a store records weekly ad spend and sales, both in thousands of dollars.
| Week | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
|---|---|---|---|---|---|---|---|---|
| Spend ($k) | 10 | 14 | 8 | 16 | 12 | 20 | 6 | 18 |
| Sales ($k) | 82 | 90 | 75 | 99 | 85 | 112 | 70 | 103 |
A one-variable least-squares regression, sales = a + b × spend, gives:
- Mean spend = 104 ÷ 8 = 13.0; mean sales = 716 ÷ 8 = 89.5.
- Σ(spend − 13)(sales − 89.5) = 490; Σ(spend − 13)² = 168.
- Slope b = 490 ÷ 168 = 2.92: each extra $1k of spend goes with about $2.92k of sales.
- Intercept a = 89.5 − 2.92 × 13 = 51.58: the implied baseline is about $51.6k a week.
- Over the 8 weeks, baseline = 51.58 × 8 = 412.7, media = 2.92 × 104 = 303.3, so media explains 303.3 ÷ 716 = 42% of sales.
The fit looks excellent (R² = 0.99), and that is exactly the trap. Eight points with one variable is below even the 10-to-1 rule, there is no adstock or saturation, and nothing controls for the fact that the high-spend weeks may also have been promotion weeks. The slope is a hypothesis about marginal return, not a budget instruction. To trust it you would need two years of weeks, controls and ideally a holdout test.
When MMM is not worth it yet
Use this rule of thumb before you commission or build a model:
- If you have less than about 18–24 months of weekly data, or most of your spend is in one or two digital channels, then start with a spend table and attribution: import spend by channel, compare cost per acquisition and ROAS, and read several attribution models side by side.
- If a meaningful share of budget goes to channels with no click (TV, radio, out-of-home, podcasts), then MMM is the main way to measure them and the history requirement is worth meeting.
- If MMM and attribution disagree on a channel by more than you can explain, then run an incrementality test on that channel before moving budget.
Common mistakes
- Treating a short regression as an MMM. Without adstock, saturation and controls, a spend-vs-sales slope mostly measures what else happened in the high-spend weeks.
- Never varying spend. A channel held flat for two years cannot be estimated, no matter how good the software.
- Reading the coefficient as average ROAS. The response curve is about marginal return at current spend; the next dollar usually returns less than the average.
- Mixing outcome definitions. Changing what counts as a conversion halfway through the history breaks the model.
- Using MMM for daily optimisation. It is a planning tool refreshed monthly or quarterly; day-to-day bidding needs faster signals.
MMM inputs and checks in TapCub
TapCub does not build media mix models. What it gives you is the spend-plus-attribution view the rule above recommends, and clean weekly series to feed a model built elsewhere.
- Bring spend next to outcomes. In All-platform analytics › Business › Channel ROI, import daily spend by channel, source and campaign (CSV columns
day,channel,source,campaign,cost, plus optional currency, clicks and impressions). The page shows cost per click (spend ÷ sessions), cost per acquisition (spend ÷ purchases on the order ledger) and return on ad spend (revenue ÷ spend), grouped by channel, source or campaign. Re-importing the same day, channel, source and campaign overwrites the row, so you can correct spend safely. - Read attribution models side by side. In All-platform analytics › Behavior › Behavior analysis, choose the attribution view and tick first, last, linear, position and time decay together; each model becomes a column, so a channel that only assists stands out. The channel ROI spend table and the attribution model are separate calculations, so compare them rather than expecting one to equal the other. How to compare attribution models walks through the settings.
- Export weekly series for a model. In All-platform analytics › Behavior › Event analysis, set granularity to week and chart sessions by channel and the property sum of
revenueon purchases, then export the table as CSV. Bots and AI crawlers are counted separately in TapCub, so they do not inflate the sessions column you hand to a model. - Mark the controls. Add promotions, price changes and launches in Configuration › Annotations, so the weeks that need a control variable are visible on the chart and easy to list.
Sources
- Meta Open Source, Robyn: Analyst's guide to MMM.
- Google Developers, About Meridian.
- Jin, Wang, Sun, Chan, Koehler (Google), Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects, 2017.
- Chan, Perry (Google), Challenges and Opportunities in Media Mix Modeling, 2017.
Was this explanation clear?
We use your answers to fix and expand this page.
Couldn’t send. Please try again in a minute.
Thanks for the feedback. We’ll use it to improve this page.You’ve already sent feedback on this page. Thanks.

Data Analyst
Explores traffic sources, conversion funnels, and retention, using clear metric definitions and concrete examples to explain what the numbers mean.