Marketing mix modeling: foundations & core concepts

CocaCocaCustomer Acquisition Cost: total sales and marketing spend divided by the number of new customers acquired over the same period.View full definition →-Cola spends on the order of $4 to $5 billion a year on advertising, across roughly 200 markets, in a category where most purchases happen at a till with no login, no cookie and no receipt that links back to anything. No user-level tracking system will ever connect a television spot in Mexico City to the bottle sold three days later in a corner shop. Somebody still has to decide how that money splits between TV, outdoor, digital video, price promotion and sponsorship. Marketing mix modeling is the answer the industry settled on decades before click tracking existed, and it came back into fashion because click tracking has been coming apart since 2021.

What marketing mix modelingmarketing mix modelingA statistical approach that estimates how each marketing channel and other factors drive sales, guiding budget allocation.View full definition → actually is

Marketing mix modeling (MMM) is a statistical technique, a form of multivariate regression, that explains one aggregate business outcome (weekly sales, revenue, units, subscriptions) as a function of everything that could plausibly have moved it. You feed it two or three years of history at weekly or monthly grain: spend by channel, price, promotional depth, distribution, competitor activity, seasonality, macro conditions. It returns coefficients that quantify how much each input contributed.

The unit of analysis is a time period, not a person. That one design choice is what separates MMM from attributionattributionA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.View full definition →: the model never needs to know who saw the ad.

Written out in words, a mix model says: sales in week t equals a baseline, plus the effect of each channel's spend once that spend has been adstocked and passed through a saturation curve, plus the effect of price, distribution, competition and seasonality, plus an error term. Almost everything interesting in MMM lives in those two transformations applied to raw spend before the regression ever sees it.

The headline output is a decomposition. Total sales split into a baseline (what you would have sold with zero marketing, carried by brand equitybrand equityThe commercial value your brand adds beyond functional product attributes: the price premium, preference and loyalty it generates.View full definition →, habit and shelf presence) and incremental sales attributed to each activity on top. A mature, heavily distributed brand like Coca-Cola typically shows 60-70% of volume sitting in baseline. A young performance-led brand might show 30%. Neither number is good or bad on its own, but you cannot manage what the model has not separated.

Sub-concept 1: adstock and carryover

Marketing does not stop working the week the money leaves the account. A TV campaign that ran in October still shifts purchases in December. That decay is called adstock, and the standard formulation is recursive: this week's effective advertising pressure equals this week's spend plus a decay rate times last week's effective pressure. A decay rate of 0.7 means roughly a two-week half-life; 0.9 stretches the tail out for months.

TV and video usually carry long adstock, often eight to thirteen weeks of measurable residue. Paid search carries almost none, because the ad meets intent that already existed and converts within days. A model with no adstock term systematically undervalues brand channels and flatters performance channels, which is one mechanical reason the 2015-2020 measurement stack pushed so much money into the bottom of the funnelfunnelThe customer journey from awareness to purchase, typically Awareness, Interest, Consideration, Decision, Action, with prospects narrowing at each stage.View full definition →.

Sub-concept 2: saturation curves

Spend and response are not proportional. The curve is concave, sometimes S-shaped with a slow start below a threshold of visibility, and it always flattens. The first $500K of TV in a market may generate strong lift, the second $500K noticeably less, the third almost nothing beyond frequency waste.

This is where average ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → and marginal ROI part company, and the distinction matters more than any single number in the model. A channel can report a healthy blended ROI of 3.0 while the next dollar into it returns 0.4, because you are sitting on the flat part of its curve. MMM estimates where each channel currently sits, which is the only way to answer the question finance actually asks: if I give you 15% more, where does it go and what comes back?

Sub-concept 3: base versus incremental sales

The baseline is the accumulated residue of years of brand building, distribution and consumer habit. It moves slowly, which is exactly why it is worth watching. Binet and Field's analysis of roughly a thousand IPA case studies found that brands putting around 60% of budget into long-term brand building outperformed activation-heavy peers on profit growth, and the mechanism is visible in a mix model: activation borrows from the baseline, brand building feeds it.

A baseline eroding across four consecutive quarters is the most serious signal MMM produces. It means the structural advantage is draining, and no amount of promotional spend refills it.

Sub-concept 4: exogenous variables and control factors

A usable MMM controls for everything that moves sales and is not your marketing: competitor pricing and promotions, consumer confidence, distribution gains and losses, and weather where the category is sensitive to it. If your ice cream brand had a record summer and temperaturetemperatureA setting that controls how random or predictable an AI model's output is: low keeps it safe and consistent, high makes it more varied and creative.View full definition → is missing from the model, the heat gets credited to whichever campaign happened to be live. That is not a rounding error. It is a decision to spend next year's budget on something that did nothing.

Sub-concept 5: why aggregate econometrics survived user-level tracking

Two reasons, one structural and one recent.

The structural one: user-level measurement can only see the channels that carry an identifier. It cannot price in a shelf listing, a 30-second spot, a stadium board, a competitor's discount or a price increase, and it cannot observe the counterfactual, the sale that would have happened anyway. Attribution allocates credit inside trackable digital. MMM estimates incremental contribution across the whole P&L, at the level budgets are actually set.

The recent one: Apple's App Tracking Transparency, launched with iOS 14.5 in April 2021, made cross-app tracking opt-in, and most users declined. Meta told investors in February 2022 that the change would cost it roughly $10 billion in revenue that year, and the damage to advertiser-side measurement was of the same order. Google spent years planning to remove third-party cookies from Chrome, then shelved the plan, but the signal loss from browser and platform restrictions is permanent enough that the platforms themselves now ship modeling tools: Meta open-sourced its Robyn framework in 2021, Google released Meridian more recently. Both are useful. Both come from companies that sell the media the model is scoring, so treat their default priors as a starting point rather than a verdict.

Real-world cases

Case 1: Coca-Cola in 2020. Revenue fell around 11% as restaurants, stadiums and cinemas closed, while the company also cut marketing spend hard. Those two facts are easy to conflate and the conflation is expensive. A mix model separates them: away-from-home availability is a distribution variable, media pressure is another, and the pandemic shock sits in a control term. Without those controls, a naive model reads the sales collapse as proof the marketing cut destroyed demand, and the following year's budget gets set on a false reading of the same data.

Case 2: Meta after ATT. The company that had spent a decade telling advertisers that user-level attribution was the modern standard began publishing an open-source econometric package once the identifiers thinned out. Robyn does the things described above: adstock transformation, saturation curves via Hill functions, decomposition against a baseline. The interesting part for a CMO is the admission embedded in the release. The most tracked medium in history reverted to weekly aggregate regression, which tells you how much of the promise of person-level measurement rested on infrastructure that no longer holds.

CMO action items

• Audit your data before commissioning anything. You need two to three years of weekly spend by channel, plus price, distribution, promotional calendar and competitive spend. Missing history is the single most common reason a model comes back with wide confidence intervals and no usable answer.
• Ask for response curves and marginal ROI by channel, not a table of channel ROIs. A ranked ROI list tells you nothing about where the next dollar should go.
• Track the baseline quarter over quarter as a brand health indicator. Growing baseline means past investment is compounding; a sustained decline means the next promotion is borrowing from a shrinking account.

Common mistakes that kill results

Mistake 1: running MMM once and treating the output as permanent. Competitors move, prices change, channels mature. A model fitted on 2021 data misprices 2024 media. Refresh annually at minimum, quarterly on a rolling window if the category moves fast.

Mistake 2: asking too much of too little data. Three years of weekly observations gives you about 156 data points. Load 40 variables into that and the coefficients become unstable, especially when channels move together (everyone raises spend in Q4), which is textbook multicollinearity. Fewer, better-specified variables beat a model that includes everything.

Mistake 3: reading correlation as proof. MMM finds statistical association in observed history; it does not run an experiment. The model earns its authority only when its estimates are checked against controlled tests, and that calibration work, along with the politics of who gets to build the model, belongs to the later rungs of this module.