Marketing mix modeling for DTC brands: when is MMM worth it?

Marketing mix modeling is worth it for a DTC brand once you have clean weekly history, weeks where spend on each channel actually moved, and at least one test to check the model against. Before that, a model mostly turns messy data into confident-looking curves. For TV, podcasts and other channels with no clicks, the order that works is facts, then tests, then the model.

Most of the work is the data underneath: weekly spend by channel, orders from your site and Amazon, and one agreed definition of a new customer. That is the DTC data warehouse and reporting work we do, and it's the part that decides whether any model is worth reading.

Key takeaways

• MMM (marketing mix modeling) is a statistical model fit on your own weekly history of spend and sales by channel. It plans forward from your past.
• It draws a curve through weeks you spent more or less. It can't see past spend levels you've never tried, so a channel you never changed has a guessed curve.
• Tests, such as switching TV off in some cities for a few weeks, are the closest thing to truth. The model fills in between tests and should be checked against them.
• Get the facts clean first: weekly spend for every channel, including TV and podcast files, orders on your site and Amazon, and one new-customer definition.
• MMM comes last, after clean history and at least one test. It's still just a model.

What is marketing mix modeling?

Marketing mix modeling is a statistical model that estimates how much each channel adds to sales, using your own weekly history of spend and sales. It looks at weeks when you spent more on TV, or less on Meta, and asks how sales moved in those weeks.

The output is a curve for each channel, showing what a given level of weekly spend tends to bring. Because it works from totals, not clicks, it covers TV, podcasts and radio. It is your own history used to plan forward, which is both its strength and its limit.

Why can't MMM see past what you've tried?

MMM can't see past what you've tried because a curve is only drawn where there are data points. If you've spent between $40,000 and $55,000 a month on TV for two years, the model knows something about that range. It knows almost nothing about $90,000 a month.

Beyond the highest spend you've actually run, the curve is guesswork until you test. The model still draws a line there, from the shape it was told to assume.

The same goes for a channel whose spend barely moved. Here is how confidence looks for three channels in a synthetic example, where the next $10,000 a month is the question:

"Back per $1" here is first-year profit per dollar of ad spend. The TV figure might be right. With spend that never moved, there's no way to tell from history alone. For more on reading these curves, see average vs marginal return on ad spend.

What does each measurement method answer?

Each method answers a different question. Attribution says who clicked, surveys say what customers remember, tests say what actually changed, and MMM says where the next dollar likely goes.

Attribution tools still have a place. A tool like Triple Whale can feed the warehouse as one more source, next to surveys and test results. See how to bring Triple Whale attribution into a data warehouse. None of them replaces the agreed rule for who gets credit.

Why do tests come before the model?

Tests come before the model because a test tells you what's true, and a model can only estimate it. Incrementality is the number of sales a channel actually caused, the ones that wouldn't have happened without it. A test measures it directly.

A geo holdout test is the usual one for TV. You switch TV off in some cities for a few weeks, keep it on in similar cities, and measure the drop in new customers on your site and on Amazon. The gap between the two groups is what TV was doing.

Measure Amazon, not just your site. TV often sends people to search for you on Amazon, so a site-only test can undercount it. How to apportion TV spend to Amazon CAC shows how to turn that split into a fair cost per new customer for each channel.

What should you have in place before MMM?

Before MMM, you need clean history, real variation and a test to check against. Here is the readiness checklist we use:

Write the definitions down before any modeling starts. Why AI gives different answers to the same metric question shows what happens to a "new customer" count when three teams each use their own.

If you can't tick all five, the model isn't the next step. The data is.

What is the right order: facts, tests, then the model?

The right order is facts first, then tests, then the model, because each step depends on the one before it. A test needs clean facts to measure against. A model needs both clean facts and tests to be worth trusting.

Step 1, now: learn from the weeks you already have

Your past weeks already hold some answers. Weeks where you spent more or less on a channel show where its curve starts to bend. Line up weekly spend and weekly new customers by channel, and you'll often see which channels have room and which are flat.

Step 2, next: run spend tests

Pick the channel you're least sure about and the one with the most money at stake. For TV, run a geo holdout. For a channel that looks underfunded, run a 4-week push. Each test answers one question for certain.

Step 3, later: MMM on every channel at once

Once history is clean and you have tests to check against, fit an MMM on every channel at once. Check its answers against your tests. Where it disagrees with a test, trust the test and find out why.

It's still just a model. Re-fit it as new weeks come in, and re-test when spend, creative or the channel mix changes enough to move the answer.

Which MMM tools can a DTC brand use?

There are free, open-source MMM tools from the big ad platforms, and many paid ones. Two well-known open-source options are Google's Meridian and Meta's Robyn.

Both are code libraries, not dashboards. Someone on your side needs to be comfortable with code and statistics to run them, and to explain the results to a CFO in plain words.

The tool matters less than what goes in. Fit on patchy data with no tests, either one gives confident curves with nothing behind them.

How much can a better budget split be worth?

Moving even a small share of budget to where the next dollar works harder can be worth a lot. This is an illustration, not a client result.

Say a $50M DTC brand spends $12.5M a year on ads. If 10% of that, $1.25M, sits in a channel where the next dollar brings back $0.33, and could move to one bringing back $1.11, the difference is $0.78 × $1.25M. That is about $975K more first-year profit a year.

The real gain would be lower, because returns fall as you add spend. The point is the order of magnitude, and you only get it if the curves are right.

Frequently asked questions

What to do next

Run your channels through the readiness checklist above. If weekly spend for TV and podcasts lives in invoices and spreadsheets, start there: that's the facts step. For the bigger picture of definitions, data and dashboards, read the white paper on how to build a DTC data practice, or book a call and we'll look at which step you're on.