How We Built the Best MMM for DTC

7-FIGURE BRANDS — Tickets for our in-person workshop are now available. 7-FIGURE BRANDS — Workshop tickets now available. Learn More →

Two years ago, I thought I understood Marketing Mix Modeling for DTC. I was wrong in ways that cost us time, cost clients money, and forced me to unlearn most of what I thought I knew. This is the story of three versions of an MMM product, what broke each time, and the three things I now believe are actually true about measurement.

If you are running a 7-figure or 8-figure ecommerce brand and trying to understand which channels are actually driving your growth, this is the context I wish someone had handed me at the start.

Why MMM Felt Like the Answer

For a long time, Marketing Mix Modeling had the allure of an exclusive black box. Obscure mathematical frameworks reserved for high-six-figure engagements at the largest brands. The rest of us just had to guess. Cookie-based attribution was decaying. Pixel data was losing signal. MMM felt like the sophisticated solution that would cut through all of it.

The pitch was irresistible: give it your total spend and sales data, let the math tell you which channels were working, and get back the perfect mathematical solution to your marketing questions.

That is what I thought.

Turns out MMM is a lot more about judgment and art, aided by machine learning and Bayesian statistics. The math is cold and it matters, but math alone is not how it gets good.

Version 0: 2020-21 -- I Thought I Knew What I Was Doing

For the first version, we used Robyn, Meta's open-source MMM package. I thought just running the model was the hard part. We connected it to BigQuery and turned it into a product. What else would be required?

There was only one problem: a lot of the results made no sense.

Meta would claim unreasonable amounts of credit. Google response curves would yield coefficients drastically different from reality. The model produced answers, but it was hard to justify them. We had built something that looked like an MMM product before we really understood MMM.

"We had built something that looked like an MMM product before we really understood MMM. The math ran. The outputs appeared. But we could not explain them -- and that should have been the red flag."

The lesson from version 0 was not about the model. It was about the difference between running a tool and understanding what it is actually doing.

Version 1: Spring 2022 -- I Realized I Had No Idea What I Was Talking About

At this point, I went back to basics. I researched, hired experts, consulted professionals. I read papers on Bayesian estimation, prior distributions, time series modeling, model selection, and experiment calibration. I studied econometrics again.

We moved from Robyn to a more rigorous Bayesian model. We had built-in priors, better diagnostics, modeled adstocks and diminishing returns, and we understood how to explain the entering variables. Academically, this version was much better.

But it still had serious problems.

You had to account for holidays, tracking changes, and events that happened once and might never happen again. If you made all the right adjustments, the model would work. If you did not, it did not. There was also a scaling problem: the model ran in 48 hours, and every small change in a variable could change the answer dramatically.

If you cannot explain something in simple terms, you do not really understand it. And that was me. I had better tools and better theory, but the output still could not answer the questions a client actually needed answered.

Version 2: Summer 2022 -- It Started to Click

The pressure to rebuild came from clients not finding enough value in the output. The questions were specific: "You are recommending we move from $5k on Meta to $10k, but we are already at 70% of our revenue target -- what exactly should we do?"

That is where the real evolution happened. The new version kept the Bayesian formulation but closed the gap between the model and the business. It connected the analyst to the actual context of what was happening inside each brand.

"A statistical model is expertise moving through time. It is a flashlight pointed at a specific estimation problem -- not a black box that replaces judgment. The moment we understood that, everything changed."

DTC data has context. Notes from client meetings, data audits, tracking changes, campaign history, creative shifts. That information lives in Slack, spreadsheets, or somebody's head. And the model needs all of it. Building an MMM became more about understanding the business than understanding the model.

Three Things I Believe After Two Years of MMM

1. MMM Is an Art

The tools are getting better as science. But that means the people running the model matter more than the model itself. It takes hundreds of hours to get good at this. You have to watch real models navigate real data. You have to learn which diagnostics matter and which ones lie to you. The tool does not make you good. The reps do.

2. Most Good Commercial MMMs Produce Similar Results

Most leading commercial implementations use the same core theory: built-in seasonality, adstocks, diminishing returns, separation of marketing from other business factors. The difference is not the model. The difference is the context and the documentation. Whether the system knows about your promotions, tracking changes, and creative shifts. Whether the person interpreting the results understands how your brand actually makes money. That is the variable that matters.

3. Context Matters More Than the Model

MMM in a vacuum is a very expensive toy. Good priors come from understanding the business holistically, not from a formula. iOS 14 changed the measurement environment. ATT and IDFA introduced bias. A brand at 7-figure revenue and a brand at 9-figure revenue have completely different channel logic, conversion behavior, and policy dynamics.

The model needs clean data AND the context behind the data: what changed, what was an outlier, how the channel usually works, and what the team already knows. Without that, you are running a very sophisticated analysis on the wrong questions.

This is exactly why we built the Prophit Engine the way we did. The math is there. The Bayesian rigor is there. But the system is designed to incorporate the business context that makes the math trustworthy -- the notes, the audits, the creative history, the channel knowledge that lives with the people running your brand every day.

What This Means for Your Brand

If you are evaluating MMM for your ecommerce brand, here is the honest checklist: find out who is actually running the model and how long they have been doing it. Ask what business context they capture alongside the data. Ask how they handle tracking changes and promotional events. Ask how fast they can turn around an answer when your question changes. The answers to those questions matter more than the name of the model they use.

Measurement is not a product you buy. It is a practice you build.

Frequently Asked Questions

What is Marketing Mix Modeling for DTC and how is it different from attribution?

Marketing Mix Modeling for DTC uses aggregate spend and sales data over time to estimate channel contribution, rather than relying on cookies or pixel-level tracking. Unlike last-click or even multi-touch attribution, MMM is not affected by iOS privacy changes or IDFA deprecation. It models the relationship between your media investments and revenue using econometrics and Bayesian statistics, making it particularly valuable for understanding channels that are hard to track directly, like TV, influencer, or brand spend.

How much data does a brand need to run MMM reliably?

Generally, you need at least 52 weeks of consistent spend and revenue data across the channels you want to model. More history gives the model more signal, especially around seasonality and channel interactions. Brands with fewer than 18 months of data can still run MMM, but the results require more careful prior setting and tighter calibration against known experiments or lift tests. The model is only as trustworthy as the data going into it.

Why do so many MMM implementations produce outputs that do not match reality?

The most common failure is missing context. A model that does not know about a major promotional event, a tracking outage, a significant creative pivot, or a platform policy change will try to explain those anomalies through the channels it can see. That produces inflated or deflated coefficients that look mathematically valid but are wrong. The second most common failure is using a model built for large CPG brands on a DTC business with completely different conversion dynamics and channel mix. The math can run. That does not mean the answer is right.

How does CTC's Prophit Engine incorporate MMM into brand decisions?

The Prophit Engine is designed around the insight that context makes models trustworthy. It combines Bayesian MMM with the operational knowledge of the team managing your brand: creative history, channel notes, tracking audits, and promotional calendars. The result is a measurement system that can answer specific business questions, not just produce coefficients. We designed it this way because we spent two years learning what breaks when you leave the business context out.

Ready to Build a Measurement System That Actually Works?

We spent two years breaking and rebuilding our approach to MMM so you do not have to. If you are a 7-figure or 8-figure ecommerce brand trying to get clearer on what is actually driving your growth, we built the Prophit Engine for exactly that.

Talk to Us

Taylor Holiday is the CEO of Common Thread Collective. A former professional baseball player who lucked into entrepreneurship over a decade ago, Taylor lives in Southern California with his amazing wife and three kids — “who are my world.” He’d love to connect with you on Twitter or LinkedIn.