Mix modelling for smaller budgets

Mix modelling is a statistical method that relates business outcomes to marketing inputs over time. It was once the preserve of large advertisers because it required expensive consultancies. Open tooling has changed that.

Why it matters now

It uses aggregate data only: weekly spend per channel, weekly sales, and control variables such as season, weather, pricing and competitor activity. No cookies, no consent, no user-level tracking.

That makes it the one measurement approach unaffected by tracking loss, and the only practical way to measure print, radio, outdoor and sponsorship alongside digital channels in the same framework.

What it requires

The critical requirement is variation. If a channel ran at the same spend every week for three years, the model cannot separate its effect from the baseline.

That has a practical implication worth acting on now: vary spend deliberately, including occasional pauses, so future analysis has something to work with.

What it tells you and what it cannot

It answers: how much of last year's revenue is attributable to each channel, what the diminishing returns curve looks like, how long each channel's effect persists, and what the optimal reallocation would be.

It cannot answer: which individual customer came from where, what happened this week, or whether a specific creative worked.

It is a strategic instrument, reviewed quarterly, not an operational dashboard.

The realistic accuracy

Confidence intervals in a mid-sized model are wide. A result reading "radio contributed between 4 and 11 per cent of enquiries" is typical.

That is still far more information than an attribution tool that reports zero for radio because it cannot see it.

Getting started without a specialist

1. Build the data set first: one spreadsheet, one row per week, one column per input. This is 80 per cent of the work and can be done by anyone.
2. Keep it updated weekly from now on, regardless of whether you model yet.
3. After two years, run a first model with open tooling or a few days of contracted analysis.
4. Validate the model's suggestion with a holdout test before acting on it at scale.

Step four matters. A model that recommends doubling a channel should be tested in one region first.