When not to use MMM
MMM is the wrong purchase when spend is concentrated in one or two channels, history is under a year, spend never varies, or the budget question is binary rather than allocative. In those cases a geo experiment answers your actual question for a fraction of the cost, and this page exists because no vendor will say so.
By Oliver Wakefield-Smith · Updated 2 August 2026 · Every number on this page resolves to a primary source in the ledger below.
The three disqualifiers
Too few channels. With one or two channels, there is no mix to model; the question is "does this channel work", which is an experiment's question. A model's comparative advantage begins where simultaneous channels make experiments combinatorially impossible, in practice around four or more.
Too little history. Under roughly a year of weekly data, seasonality and marketing are statistically inseparable; under two, barely so.1 No framework option and no vendor fee changes the information content of 40 rows. The fix costs nothing but patience: log data properly and run experiments while history accumulates.
No variation. If every channel has spent the same weekly amount for two years, the data cannot distinguish those channels' effects from baseline. Robyn's docs are blunt that the search algorithm needs signal to find.2 Either introduce deliberate spend pulses first, or accept that the model will return expensive shrugs.
What a too-small brand should do instead
The unglamorous stack: platform dashboards for tactics; one or two geo-lift tests a year on the biggest question (a testing platform like Haus, or a hand-rolled matched-market test, both work);3 and a simple spreadsheet triangulation of blended CAC against experiment results. That combination costs a fraction of any MMM contract and answers the questions a two-channel brand actually has. Graduate to MMM when the readiness checklist says so.
When a running MMM stops earning its keep
Retire or pause the model when its inputs stopped resembling the business (category pivot, 3x growth, channel mix overhaul), when refresh-to-refresh estimates swing too wide to act on, or when nobody has moved a dollar based on it in two quarters. A model that does not change decisions is a subscription to a chart. The honest failure story vendors omit: most abandoned MMMs did not fail statistically; they failed organizationally, unread, unrefreshed, uncalibrated. Budget the attention before the software.
Ready anyway? Return to build vs buy with clearer eyes.
Source ledger
Every numeric claim on this page resolves to one of these primary sources.
- [1]Google Meridian documentationhttps://developers.google.com/meridianretrieved 2026-08-02
- [2]Meta Robyn feature documentationhttps://facebookexperimental.github.io/Robyn/docs/featuresretrieved 2026-08-02
- [3]Haus public sitehttps://www.haus.io/retrieved 2026-08-02