Marketing Mix Modelling Basics for Small Teams

Click-based attribution struggles with offline sales, TV, hoardings, influencers and users who decline tracking. Marketing mix modelling (MMM) takes a different route. It looks at aggregate data over time and asks which spending patterns moved results.

The core idea

MMM is a statistical approach, often a regression, that relates a business outcome (sales, leads, orders) to inputs over time. Inputs include spend per channel, pricing, promotions, seasonality, festivals and sometimes weather or competitor activity.

If sales consistently rise a few weeks after radio spend increases, and the pattern holds after accounting for festivals and discounts, the model assigns some of the lift to radio. No individual user is tracked.

What data you need

• Weekly (or daily) sales or leads, ideally two or more years of history so seasonality is visible.
• Spend and activity by channel for the same periods.
• Prices, discounts and stock issues that affected demand.
• Dates of festivals, sales events and major external shocks.
• Notes on anything unusual, such as a website outage.

Clean, consistent data matters more than a fancy model. See our notes on cleaning marketing data.

Two ideas worth knowing

• Adstock (carry-over): advertising keeps working after it runs. A model usually lets effects decay gradually.
• Diminishing returns: doubling spend rarely doubles results. Models estimate a saturation curve per channel.

These make the output more realistic than a simple "each rupee gives X" rule.

What you get out of it

• An estimate of each channel's contribution to results.
• Response curves that suggest where extra budget helps less.
• Scenario comparisons, such as shifting budget between channels.

Treat these as informed estimates with uncertainty, not exact facts.

A hypothetical example

A furniture brand sells online and in showrooms. Click reports credit search and Instagram, while showroom visits rise around newspaper campaigns. An MMM on three years of weekly data suggests newspaper spend contributes to showroom-driven sales and that search shows diminishing returns at the highest budget levels. The team runs a regional budget test to validate before moving large amounts.

Limits to respect

1. Needs enough variation: if you spend the same amount every week, the model cannot separate channels.
2. Correlation is not proof, especially when channels are scaled together during festivals.
3. It works best at broader levels, not for tweaking individual ads.
4. Results depend on assumptions, so ask how they were chosen.

How to start practically

1. Collect and tidy a weekly dataset.
2. Try open-source tools or ask an analyst or agency with experience.
3. Validate the findings with geo tests or holdouts, as discussed in incrementality and testing.
4. Combine it with platform data and your attribution models instead of replacing them.