Marketing Mix Modeling Reality Check for Teams Under $10M Revenue

A founder once showed me an MMM vendor proposal for a company doing about $3M a year across four channels. The deck was gorgeous. The underlying math was never going to work, and nobody in the sales process had said so.

The statistics problem nobody mentions

Marketing mix modeling needs enough weekly variation in spend across enough channels over enough history to statistically separate each channel's contribution from the others and from seasonality. That's typically 2+ years of weekly data with real spend fluctuation, not a steady autopilot budget. A company spending $15k a month split evenly across Meta and Google, flat for a year, has handed the model almost nothing to learn from — the spend barely moves, so the model can't tell you what moving it would do.

MMM doesn't fail quietly at small scale. It produces a confident-looking number with no statistical basis behind it, which is worse than no number.

What I tell teams in this range instead

• Run real holdout tests on your two biggest channels — geo holdouts or simple on/off periods give you an actual causal number, not a modeled approximation.
• Use last-click or a simple multi-touch model as a directional guide, not a precision instrument — it's honest about its own limitations.
• Deliberately vary spend for a quarter if you want MMM to eventually be viable — the model needs the variance you're currently avoiding.
• Revisit MMM when you cross roughly $10-15M with five-plus channels and real budget flexibility, where there's enough signal to separate from noise.

I'm not against MMM. I'm against selling a statistical method to companies whose data can't support what the method needs to work. A $3M company doesn't need a better model. It needs to run a holdout test on the one channel it's unsure about and get a real answer in six weeks instead of a confident-looking answer in six months.