MMM data requirements

History: why ~104 weeks is the common floor

The model must separate marketing effects from seasonality, and seasonality only reveals itself by repeating. One year of data shows each season once; nothing distinguishes "December lifted us" from "that December campaign lifted us". Two annual cycles, about 104 weeks, is the practical floor at which those explanations separate; more is better, subject to your business still resembling the one in the data. If you changed category or tripled in size, old weeks carry less information than their count suggests.

Granularity: weekly vs daily, national vs geo

Weekly is the standard row. Daily data multiplies rows but adds noise (day-of-week effects, reporting lags) and most frameworks aggregate it anyway. The higher-value axis is geographic: revenue and spend split by region turns 104 national observations into thousands of region-weeks. Meridian is built around exactly this structure, and its documentation is explicit that geo-level data is preferred over national aggregates.¹

Channel minimums

A channel is modelable when the data contains information about it, which in practice means: enough spend share to plausibly move revenue, and enough variation to observe the moving. A channel that has never changed spend level teaches the regression nothing; its effect is statistically indistinguishable from base. Robyn's documentation makes the same point from the other side: hyperparameter search cannot rescue a flat input.² Tiny channels (a few percent of spend) come back with intervals wider than their estimates; group them or accept the honesty.

The dependent variable

Pick one revenue definition and defend it: net revenue from the finance system, orders, or new-customer revenue, depending on what marketing is for. Platform-reported revenue is disqualified; it is an output of the systems being audited. If finance and the model disagree about what revenue was, the model loses in every meeting that matters.

The controls

Price changes, promotion calendar, distribution or availability shifts, PR spikes, competitor launches if observable, and platform outages. Uncontrolled events do not politely absent themselves from the model; they get attributed to whichever channel co-varied with them. An afternoon assembling the promo calendar buys more accuracy than a month of prior tuning.

Turn this into six yes/no answers on the readiness checklist, or read the failure modes first in MMM limitations.