MMM vs attribution: what marketing mix modelling tells you that last-click never will

Short answer: attribution tells you which touchpoint happened to be nearest a sale. Marketing mix modelling estimates what actually caused it, including channels you can't track and factors that have nothing to do with marketing.

A disclosure first: I co-founded Stitch Predict, a marketing mix modelling platform, so I have skin in this game. The principles below apply whichever tool you use.

Why attribution stopped being enough

Last-click and multi-touch attribution served marketers well when most journeys happened on trackable screens. That world is gone. Privacy regulation and platform changes mean a growing share of journeys simply aren't visible, and even when they are, attribution can't tell you whether an ad caused a sale or just happened to be there when someone who was already going to buy clicked through.

That second problem, incrementality, is the expensive one. Budgets drift towards channels that are good at being last, like branded search and retargeting, and away from channels that create demand but rarely get the click.

What MMM is

Marketing mix modelling is statistical modelling that estimates how much each channel, and each outside factor, contributes to sales or another business outcome. It works on aggregated historical data, typically weekly spend and results over two to three years, rather than tracking individual people. That makes it cookieless by design and privacy-safe.

What it tells you that attribution can't

• Incremental contribution by channel. Not who got the last click, but how many sales each channel genuinely added.
• Diminishing returns. The point where the next dollar in a channel stops paying back.
• Offline and hard-to-track channels. TV, radio, outdoor, sponsorship and PR sit in the same model as digital.
• The factors you don't control. Seasonality, pricing, promotions, competitor activity and the economy, separated out so marketing isn't blamed or credited for them.
• Budget scenarios. What happens if you move 20% from one channel to another, before you spend it.

Category matters too. Research by Associate Professor Felipe Thomaz at Oxford, which I wrote about for the Marketing Association in 2024, found that a channel's influence varies enormously by category. TV had a 2% chance of influencing an auto customer but a 50% chance in personal care. Generic benchmarks won't tell you which world you're in. Your own model will.

What it doesn't do

MMM isn't built for creative-level decisions or daily bid changes. It works at the level of channels and weeks. The strongest measurement setups combine MMM for strategic allocation with controlled experiments for specific questions, and use each to check the other.

Ask this before you trust a model

The question every CMO should ask is simple: can the model predict outcomes it hasn't seen? A proper holdout test withholds a slice of data, then checks whether the model's predictions for that period match what actually happened. If a model fails that test, its recommendations are guesses with a confidence interval attached.

Other questions worth asking:

• How often is the model refreshed with new data?
• Has it been validated against real experiments?
• How does it handle channels with very little spend history?
• Does it rely on cookies or user-level tracking at all?

What it takes in practice

With Stitch Predict, models typically take six to ten weeks to deploy, need two to three years of weekly data, and run with a variance of plus or minus six to ten percent. You'll need weekly spend by channel, your outcome metric, and records of pricing, promotions and anything else that moves demand.

If you've never seen your marketing measured this way, the first model can be an eye-opener. If you'd like to talk it through, get in touch.