What is marketing mix modeling?
Marketing mix modeling (MMM) is a statistical technique that estimates the causal contribution of each marketing channel to revenue. It uses aggregate spend and aggregate revenue over time (typically 18 to 24 months of weekly data) alongside seasonality, pricing, promotions, and external variables, and produces an estimate of how much revenue each channel actually caused.
The reason MMM exists is that user-level tracking cannot see everything. Television, out-of-home advertising, podcast sponsorships, and brand campaigns rarely produce tracked clicks. MMM captures these invisible effects through their statistical relationship with total revenue. Combined with the user-level view from multi-touch attribution, MMM completes the measurement picture.
For the broader attribution context, start with what is marketing attribution. For the user-level counterpart, see multi-touch attribution.
MMM vs MTA: complementary, not competing
Multi-touch attribution and marketing mix modeling answer different questions. Running only one leaves half the picture missing.
| Dimension | MTA | MMM |
|---|---|---|
| Unit of analysis | Individual customer journey | Aggregate time series |
| Grain | Campaign, keyword, ad | Channel, portfolio |
| Cadence | Daily / weekly | Quarterly |
| Measures | Attributed revenue | Incremental contribution |
| Captures | Tracked online channels | All channels, online and offline |
| Depends on tracking | Yes (first-party) | No (aggregate spend + revenue) |
| Answers | Which channels moved this customer? | What would revenue be without this channel? |
MTA is the daily operational metric; MMM is the strategic one. The two rarely disagree about which channels are best, but they often disagree about how much budget each deserves. Those disagreements are where strategic planning actually happens.
How MMM works
At its core, MMM is a regression. You have a time series of weekly or monthly revenue, and a set of independent variables: spend per channel, seasonality, price, promotions, macroeconomic signals, competitor activity where available. The model fits a mathematical relationship between the inputs and the revenue, and the estimated coefficient for each channel is that channel's contribution.
The complications sit in the details. Marketing does not behave linearly. A dollar spent on a saturated channel produces less revenue than a dollar spent on an under-invested one. A television ad seen today continues to drive revenue for weeks afterwards. Brand building has slow, accumulating effects that a naive regression cannot capture. Modern MMM handles these through transformations (adstock for delayed effects, saturation curves for diminishing returns) applied before the regression is fit.
The output is not a single number. It is a channel-level contribution estimate with a confidence interval, response curves showing how revenue responds to spend at different levels, and diagnostics that reveal whether the model fit the data well enough to trust.
Bayesian, Ridge, and Ensemble MMM
Every MMM implementation has to decide how to fit the regression. The three standard approaches each make different trade-offs, and Attriqs runs all three in parallel so you can compare outputs rather than being locked into one method.
Bayesian MMM
Incorporates prior beliefs about channel behaviour. If you know that TV takes weeks to build impact and has a long tail, you can encode that as a prior on the adstock parameter. The output is a credible interval around every estimate, which is intellectually honest and finance-friendly. The trade-off is computational cost (Bayesian sampling takes minutes, not seconds) and the need to specify priors reasonably.
Best for: teams that want uncertainty quantification, have some prior knowledge of channel behaviour, and can invest in model tuning.
Ridge MMM
Uses regularised regression to produce stable point estimates without requiring priors. Handles correlated channel spend (when you scale Meta and Google together, for instance) more cleanly than ordinary least squares. Fast to fit, easy to re-run, and does not require statistical expertise to operate. The trade-off is that it produces point estimates rather than credible intervals, so uncertainty has to be quantified through resampling.
Best for: teams that want quick, repeatable outputs and do not have strong priors to encode.
Ensemble MMM
Combines multiple models (Bayesian plus Ridge plus variants) and weights each by model fit. The final estimate is less sensitive to any single method's assumptions. Ensemble is the default for teams that want robustness and are willing to spend more on compute; it often produces tighter estimates than either approach alone.
Best for: teams optimising for robustness and willing to accept slightly higher complexity.
Running all three in parallel is the mature pattern. Attriqs' MMM module runs Bayesian, Ridge, and Ensemble on the same data and shows the comparison in one view, with model-fit diagnostics for each.
Response curves and diminishing returns
A response curve shows how revenue responds to spend at different levels for each channel. The curve is typically concave: the first dollars on a channel produce high marginal returns, and each additional dollar produces less as the channel saturates.
The shape of the curve answers two strategic questions. Where is the channel right now (still climbing, at the knee, flattening)? And where is the budget best allocated (which channels have headroom, which are plateaued)? Both questions are invisible to MTA, which can only tell you how credit was distributed, not how revenue would change at different spend levels.
Early curve
Channel is new or under-invested. Marginal revenue per dollar is high. Scaling up produces proportional returns.
Knee of the curve
Scaling is starting to produce less per dollar. The channel is approaching efficient capacity. The decision point.
Plateau
Additional spend produces near-zero incremental revenue. Often where heavily-funded channels sit after years of scaling. Time to reallocate.
Adstock, saturation, and the math behind MMM
Two transformations make the difference between a naive MMM regression and one worth trusting.
Adstock (delayed effects)
A television ad seen today does not only drive revenue this week. It continues to build brand recall for weeks afterwards. Adstock transforms each spend week into a weighted blend of current and prior weeks' spend, so the model captures the delayed effect instead of crediting only the immediate week.
Tuning: the decay rate is channel-specific. Paid search decays quickly (days); television and brand advertising decay slowly (weeks or months).
Saturation (diminishing returns)
Doubling spend on a channel rarely doubles revenue. A saturation transform (Hill function, logistic curve, or similar) maps linear spend onto the non-linear response that reflects real-world behaviour. Without it, the regression assumes linearity and systematically over-credits saturated channels.
Tuning: saturation parameters are estimated from the data alongside contribution coefficients.
Both transformations are applied before the regression is fit. The combination of well-tuned adstock and saturation is what separates a trustworthy MMM from a spreadsheet exercise. Attriqs handles this tuning automatically as part of the model-fitting process.
Budget optimisation from MMM outputs
MMM response curves are not just diagnostic. They are the raw input for a budget optimiser that searches for the allocation that maximises forecasted revenue subject to a total budget constraint.
The optimiser solves a straightforward optimisation problem: given the response curve for each channel, what allocation of total budget across channels produces the highest forecasted revenue? The answer is a specific quantified recommendation, not a qualitative hunch. Shift $30,000 from channel A to channel B, expected revenue lift of $45,000, confidence interval X to Y.
From there, the recommendation is either reviewed manually and implemented in the ad platforms, or pushed directly through to connected platforms via the budget optimiser. This closes the loop between measurement (MMM output) and action (ad platform changes).
When to use MMM (and when it is overkill)
MMM is worth it when
- Total marketing spend exceeds roughly $1 million per year
- At least 5 to 7 channels run simultaneously
- Brand, television, podcast, or out-of-home spend is meaningful
- You have 18+ months of historical spend and revenue data
- Incremental measurement matters to the CFO or board
- iOS ATT and cookie deprecation are degrading your user-level tracking
MMM is likely overkill when
- Total spend is under $250,000 per year
- Only two or three channels run
- Less than 12 months of historical data exists
- Revenue is too variable week-to-week to model cleanly
- The business is pre-product-market-fit and spending levels are not stable
For smaller setups, MTA and careful holdout experiments often give enough signal without the MMM overhead. MMM becomes valuable once spend and channel complexity cross a threshold where aggregate causal measurement is the only honest answer.
Data requirements and limits
MMM is data-hungry at the aggregate level, but unlike MTA it does not require user-level tracking.
Required: 18 to 24 months of weekly spend per channel, weekly revenue, and basic seasonality markers (holidays, promotional periods). Spend data should reflect actual flight dates, not reported dates, for channels where these differ.
Strongly recommended: pricing history, promotional calendar, macroeconomic factors where relevant (consumer confidence index, unemployment, fuel prices for auto), and any competitor spend signals available.
Nice to have: weather, local event data, website health metrics, site speed incidents. Anything that might have affected revenue but was not marketing spend.
The fundamental limits of MMM are worth naming. It cannot measure channels that did not vary meaningfully over the modelling period; a constant spend line produces no signal. It struggles to disentangle channels that always scale together (if Meta and Google are always adjusted in lockstep, separating their contributions is hard). And it is a statistical estimate, not a physical measurement, so interval estimates matter more than point estimates.
Frequently asked questions
What is marketing mix modeling in simple terms?
Marketing mix modeling (MMM) is a statistical technique that measures the incremental contribution of each marketing channel to revenue. Instead of tracking individual customers, MMM analyses total spend and total revenue over time, along with seasonality, price, and external factors, to estimate how much each channel actually caused. It is the standard method for answering "what would revenue be without this channel?"
Is MMM the same as multi-touch attribution?
No. Multi-touch attribution (MTA) operates at the user-journey level and distributes credit for individual conversions across touchpoints. MMM operates at the aggregate level and estimates causal channel contribution using statistics. MTA is daily and granular. MMM is weekly or monthly and holistic. Mature measurement stacks run both in parallel because they answer different questions.
How much data do I need for MMM?
Typically 18 to 24 months of weekly data across at least 5 to 7 channels, with spend and revenue recorded consistently. Less data can produce directional results, but model stability improves meaningfully past the 18-month mark. Attriqs runs MMM on whatever history you have, and flags which estimates are reliable versus which should be treated as preliminary.
What is the difference between Bayesian, Ridge, and Ensemble MMM?
Bayesian MMM incorporates prior beliefs about channel behaviour (for example, that TV takes weeks to show impact) and produces credible intervals around every estimate. Ridge MMM uses regularised regression to handle correlated channel spend and produces stable point estimates without requiring priors. Ensemble combines both approaches, weighting each based on model fit, to produce outputs that are less sensitive to any single method. Attriqs runs all three.
Can MMM handle brand and upper-funnel spend?
Yes, and this is where MMM adds the most value. User-level attribution (MTA, platform pixels) cannot see the aggregate effect of brand campaigns, television, podcasts, or out-of-home advertising, because these rarely produce tracked clicks. MMM measures these by their statistical relationship to total revenue over time, which is the only honest way to defend upper-funnel budget.
How often should MMM be re-run?
Quarterly is the typical cadence for strategic planning. Major spend shifts, new channel launches, or external shocks (supply chain, pricing changes, competitive dynamics) warrant an interim re-run. MMM outputs are not daily operational metrics; they are strategic, and the re-run cadence should match the planning rhythm, not the tactical one.
Is MMM resilient to iOS ATT and cookie deprecation?
Yes, because MMM does not depend on user-level tracking. It works with aggregate spend and aggregate revenue data, which are not affected by browser privacy changes or ad platform signal loss. As user-level tracking degrades, MMM has become a standard part of every credible measurement stack for this reason.