How to Build an MMM Without Starting With the Model

The fastest way to make a Marketing Mix Model difficult to trust is to begin by choosing the model.

A common sequence feels productive: a team selects a library, imports a notebook, and starts shaping its data around the expected inputs. But every important decision has already been smuggled into the work: what outcome counts, what time period is trustworthy, which channels can be separated, which outside forces matter, and what the model is supposed to help someone decide.

A stronger MMM build begins with a decision contract and a modeling table specification. The first milestone is not a contribution chart. It is a dataset and validation plan another person can challenge before the statistical complexity arrives.

Define the decision before the outcome

“Understand marketing performance” is not a decision. It does not establish which actions the model should inform or how precise the result must be.

A bounded question looks more like this:

If the total paid-media budget remains fixed next quarter, how might reallocating spend across five controllable channels change expected weekly orders within existing channel constraints?

That question clarifies several requirements:

the outcome is orders, not a mixture of leads and revenue;
the model must support channel-level allocation;
the relevant horizon is quarterly;
scenarios must respect a fixed total and operational constraints;
weekly time may be a useful starting grain;
uncertainty must be visible because the output affects future spending.

Write the decision, intended user, refresh cadence, and constraints before collecting columns. If stakeholders disagree here, model selection will not resolve the disagreement.

Choose one stable row grain

An MMM modeling table usually has one row for each unit of time and, when the data supports it, geography. Weekly national rows are common, but they are not automatically correct.

Daily data may preserve useful variation but also contain strong noise, day-of-week effects, delayed conversions, and inconsistent spend posting. Monthly data can be too coarse, leaving few observations and hiding short campaigns. Geographic panels can add identification power when media and outcomes vary by market, but only if the geographic definitions line up across sources.

Choose the grain based on:

how quickly marketing could plausibly affect the outcome;
how spend and exposure data are recorded;
conversion and revenue delays;
the number of trustworthy observations;
geographic consistency;
the cadence of budget decisions.

Once chosen, aggregate every input to that same contract. Do not casually mix a national weekly outcome with channel data whose apparent geography is based on a different allocation method.

Audit the outcome as aggressively as the media

Teams often spend most of their preparation time cleaning channel data while assuming revenue or conversions are ground truth. The outcome can contain its own structural breaks.

Document:

the system of record;
when the business definition changed;
whether cancellations or returns are included;
how offline activity is handled;
how delayed outcomes are assigned to time periods;
currency and tax treatment;
periods affected by outages, migrations, or restatements.

Plot the outcome over the full candidate history. Mark product launches, price changes, distribution changes, promotions, inventory constraints, and other events that can create breaks. A model cannot distinguish media response from an undocumented outcome redefinition.

Build media inputs around what can be learned

Spend is often the most consistent cross-channel input, but it is not always the most meaningful exposure measure. Impressions, clicks, reach, gross rating points, or delivery metrics may describe activity more directly. The right choice depends on the channel, data quality, and decision.

For each media variable, record:

source and owner;
buying currency and unit;
native time and geography grain;
missing and zero treatment;
major platform or format changes;
whether planned and delivered values differ;
how strongly the channel moves with other channels;
periods with meaningful increases, decreases, or shutdowns.

Variation is essential. If two channels always rise and fall together, observational data may not support a stable separation of their effects. More rows do not automatically solve that confounding.

Instead of hiding the issue, document it. The model may need grouped channels, stronger priors or constraints, experiments, or a more cautious decision scope.

Add controls as causal alternatives, not decorations

Seasonality and trend are rarely enough. Price, promotions, product availability, distribution, sales coverage, weather, economic conditions, competitor activity, and major events may affect the outcome.

But adding every available column can create a different problem. A control that is downstream of marketing can absorb part of the effect the model is meant to estimate. A highly flexible control set can also fit history while producing unstable channel contributions.

For every proposed control, write:

the mechanism by which it may affect the outcome;
whether marketing may affect the control;
the expected direction;
the data source and known limitations;
the sensitivity test that will show whether it changes the conclusion.

This turns control selection into a reviewable argument rather than a column-import exercise.

Reserve history for time-aware validation

An MMM that fits the historical series closely can still be poor at estimating contribution or supporting new scenarios. Validation needs more than one fit statistic.

Design the validation plan before fitting the final model:

hold out a later time window rather than randomly mixing past and future rows;
use rolling or repeated time splits when the history allows;
compare against a deliberately simple baseline;
inspect residual patterns and large misses;
test the sensitivity of channel results to priors, transformations, controls, and time windows;
compare estimates with credible experiments and known market events;
review whether signs, magnitudes, and response curves are operationally plausible.

No single test proves causality. The aim is to accumulate evidence and expose where the model is fragile.

R. Garcia’s field observation: A technical audit of one of my own monthly models found that its output invariants were stronger than its model validation. Every market-month reconciled exactly to observed outcomes, and no channel could fall below its directly observed attribution. But the implementation had no time-aware holdout or uncertainty, and its contribution calculation mixed standardized coefficients with unstandardized feature values. The arithmetic balanced while the channel allocation remained unsafe to publish. I now treat bookkeeping checks and statistical validation as two separate release gates.

Require two validation gates

The first gate asks whether the pipeline is internally coherent:

Do inputs use the governed schema, period, currency, and geography?
Do cohort definitions and maturity rules remain stable?
Do outputs reconcile to the outcome total without double counting?
Can the run be reproduced from versioned data, configuration, and code?

The second gate asks whether the modeled explanation holds up:

Does it perform on later periods it did not fit?
Are residuals structured around seasons, launches, markets, or spend changes?
Do channel estimates remain plausible when controls, carryover, saturation, and windows change?
Do shutdowns, experiments, and known business events challenge or support the result?

A model can pass either gate and fail the other. Do not release channel recommendations until it passes both.

Specify the first modeling table

The first table should be boring and auditable. Use one row per chosen time-geography unit and keep clear groups of fields:

Group	Example fields
Keys	week_start, market_id
Outcome	orders or revenue
Media	channel spend and selected exposure measures
Commercial controls	price index, promotion depth, availability
Context	holidays, weather, macro indicators, major events
Data quality	source completeness flags, break indicators

Keep raw governed inputs separate from transformed model features. Adstock, saturation, trends, and encoded seasonality belong in a reproducible transformation layer, not silently overwritten source columns.

Know when not to fit the model yet

Pause the build when:

the outcome definition changes without a reconciliation path;
the historical window is too short for the intended complexity;
major channels have little independent variation;
spend is missing during important periods;
controls are selected without a credible mechanism;
the decision requires geographic or channel detail the data cannot support;
no one can define how recommendations will be tested or monitored.

The correct next step may be a measurement repair, a controlled experiment, a narrower model, or several more months of governed data.

The actual first deliverable

Complete the MMM data-readiness worksheet and publish the modeling-table contract internally. Review it with the people who own finance, media, analytics, sales, and the outcome system.

Only then choose the baseline, transformations, estimation approach, and software. That order makes the model serve the decision and the evidence—not the other way around.

The fuller postmortem shows why exact reconciliation, a familiar estimator, and an impressive fit statistic still do not make a channel-mix result defensible.

Follow the full Build a Marketing Mix Model hub for the next stages: baseline design, adstock, saturation, Bayesian and frequentist choices, time-based validation, experiment calibration, scenario planning, and monitoring.