Key takeaways
- Pick the budget decision your marketing mix model (MMM) needs to inform before you touch any data. A model built to answer everything can end up answering nothing.
- Build or keep up an experiment practice first, or alongside the model. Most teams add an MMM on top of experimentation, not in place of it.
- Set your channel structure at the level where you actually move budgets, then limit it to what you've tested.
- Feed experiment results into the model as it runs, and treat any disagreement between the model and your tests as a signal to investigate.
Enterprise MMMs often end up as more of an artifact than a decision-making tool.
An MMM is a statistical model that links marketing spend to outcomes like sales or revenue, and it exists to guide where each dollar goes. Whether it does that well depends less on the modeling technique and more on what you feed it and how you set it up.
Here's the order we'd work in. It's our own sequencing of what Haus has published on Causal MMM, not a fixed standard.
Start with the decision, then check you're ready
Before you gather data, write down the decision the model has to inform. Good examples are "Are we overspending on branded search?" and "Where should we put extra budget for Q4?"
The alternative is the catch-all approach, where you load in as much data as possible and see what comes out. That tends to end in analysis paralysis. The cleaner path is to remember the original goal, line up the tooling and resourcing to meet it, and then focus on the business change you're after.
Three signs suggest your measurement stack is ready for an MMM, and you can read more about them in Are You Ready for An MMM?
- You're managing a cross-channel budget, and a single-channel read no longer suffices.
- You need visibility into channels you can't turn up or down by region (geo-segment), like linear TV, podcast, and influencer.
- Your team acts on experiment results. A model anchored to experiments helps most when your team trusts those experiments.
Channel count matters too. The more channels you're already testing, the more signal an MMM has to work with. One way to picture it: experiments tell you which ingredients you need, and the MMM is the recipe for combining them.
Build the experiment foundation before the model needs it
Why an MMM alone tends to stall
Traditional MMMs run into multicollinearity. When revenue, Channel A spend, and Channel B spend all rise together, the model can't tell which one caused the lift, or whether either did. Noise from seasonality, supply constraints, and consumer spending trends adds to the problem. Two MMMs can even produce the same sales forecast while disagreeing about which channels drive it.
Experiments give the model something to check against. Without them, a statistician ends up making assumptions about marketing efficiency that nothing verifies.
Here's what a usable practice looks like. If you're starting from scratch, building an incrementality practice is the groundwork. The process has five steps: define the question and scope, design the test, launch and monitor it, analyze lift, and make a decision.
Design choices involve trade-offs. A bigger holdout gives you more confident results but withholds ads from more of your audience. Poorly matched test and control groups are a common pitfall, and synthetic controls can improve accuracy. Common mistakes include misaligned key performance indicators (KPIs) like clicks instead of revenue or new customers, false precision, trusting platform metrics, and trying to do too much in-house.
One practical upside is that true incrementality testing doesn't need user-level identifiers. It uses group-level tests across geographies or time periods, so it holds up better as privacy rules change.
The harder part is often organizational. Enterprise programs tend to stall on internal resourcing, friction, and misalignment, and teams have to work in lockstep from hypothesis to design to readout. Even sophisticated multi-business-unit brands can struggle with testing velocity in-house, because it takes a dense data science team. Wayfair found that working with a testing partner meant less time on setup and more time acting on results.
Structure channels around how you decide
Planning granularity vs. tested granularity
Granularity is the level at which the model groups your marketing activity. Start by asking at what level your team makes decisions, rather than how granular the model can get. More detail isn't automatically better, either. It can spread the same evidence thinner, which means less signal per bucket and wider uncertainty.
What you've tested sets the ceiling. In Causal MMM, your planning granularity is the target and your tested granularity is the binding constraint. A blended test can't produce unblended truth. If your MMM splits Meta into Advantage+ Shopping Campaigns (ASC), prospecting, and retargeting but your test measured Meta all-up, the lift won't map back cleanly.
The same thing happens when one Google test spans Search, Performance Max (PMax), and YouTube together. This post on matching your MMM to how you plan goes deeper.
Use a practical rule for getting started. Build channels around distinct, noticeable buckets of spend. As a rule of thumb, roll anything under about 2% of your media into a broader category until it's big enough or tested enough to stand alone. For example, a typical setup might split Google into Search Brand, Search Non-Brand, PMax, and YouTube, and Meta into Prospecting and Retargeting.
Then start broad and add detail as your evidence grows. Build at the level where budgets move, calibrate with causal signal at that level, and increase granularity as you gather more tests. Until a sub-channel has its own test, Causal MMM leans on priors from comparable channels and brands to apportion broader results, then sharpens as your own experiments arrive.
Anchor the model with experiments
Treat results as ground truth
We treat an experiment result as ground truth. The experiment doesn't just serve as a foundation for the model; it anchors the response curve, which shows incremental impact at different spend levels. Tests also give you the baseline, along with the incremental result. Without a baseline, you can draw a slope between two spend levels, but you won't know where the curve should start.
Timing matters too. A more accurate approach feeds incrementality results into the model as it runs, before it generates outcomes, not after. Better still is a model that accounts for when an experiment ran and weighs it against your seasonality.
If you want the longer argument, read why incrementality testing belongs with your MMM. How it works lays out the approach.
A Bayesian MMM needs priors, a starting view of how each channel's spend relates to outcomes. They can come from academic studies, benchmarks, industry knowledge, or past MMM and attribution results, but those sources can be generic or correlation-based. For channels you haven't tested, we start from incrementality signals, either trusted third-party studies or priors tailored to your own experiments, until direct evidence shows up.
You don't need full test coverage to begin. Causal MMM uses GeoLift, Haus' geo-experiment tool, where tests exist and extends estimates through Conversion Lift studies. Testing the channels you can test also improves estimates for the ones you can't.
One question is worth asking any model you're considering: Was it built to accept experiments in the first place? Feeding a result into a curated set of priors can end up fighting the model if experiments weren't part of the build.
What implementation looks like in practice
StockX,OluKai, and Sunday Lawncare
StockX had managed its largest expense with static budgets and gut feel, with little YouTube visibility beyond last-click. A YouTube incrementality test became an input to its Causal MMM, and the model surfaced a scaling opportunity that a correlation-only model had buried as a "nice to have."
OluKai's story shows what's at stake when the model and the tests disagree. According to Haus' GeoLift to cMMM write-up, OluKai's legacy MMM favored view-through-heavy channels like TikTok, Snap, and YouTube heading into Q4, while GeoLift said TikTok and Snapchat were less incremental than Meta. After following the legacy MMM, OluKai's customer acquisition cost (CAC) came in 20% higher than predicted.
With Causal MMM, the team moved about 50% of view-through spend into Meta and cut total marketing spend about 15%. CAC dropped about 20%, forecasts matched actuals, and the gains held for several weeks. That's one brand's experience, and your results will likely differ.
Sunday's weekly rhythm shows what that looks like in practice. Sunday gets a causal read every Monday morning, based on the prior week's data. Its response curves combine GeoLift and Meta Conversion Lift, and Causal MMM weights experiments dynamically so older or out-of-season tests stay representative. This case study covers Sunday's setup in more detail.
Avoid the build-it-yourself trap and keep the model trusted
An in-house MMM looks easy at first. You take an open-source model, add two data scientists and three engineers, and you can have response curves in weeks.
What follows is often predictable. Month one is a honeymoon. In months 2–4, discrepancies show up between multitouch attribution (MTA), MMM, and experiments. In months 4–6, teams fall into the prior-tuning trap, and in months 6–12, credibility erodes.
Prior tuning is essentially cherry-picking, because you're choosing how the model starts to get results you want to see. Around month 12, you face a choice between investing heavily in infrastructure and staff or paying an outside provider, since keeping a model current is a continuous investment. This post on building an in-house MMM covers the trade-offs, and some teams choose a measurement platform instead.
Here's a simple acceptance test we'd suggest, though it's our own framing rather than a published standard: Before you rely on the model, check whether its recommendations line up with the experiments you already trust. If they agree, move fast. If they don't, investigate. A model that tells you to scale a channel your tests show isn't incremental creates a trust problem.
Ben Bolognini, OluKai's VP of Ecommerce, puts the tiebreaker plainly in the same GeoLift to cMMM post: "When clear experimentation and directional measurement results disagree, incrementality should win."
Then keep your tests fresh. Platforms change targeting, accounts get restructured, and a test can fall outside the model's training window. There's no fixed cadence, but more frequent channel-level re-runs plus seasonal adjustment can help keep results trustworthy. A test run in high season may not apply off-season, either.
Put the model to work on a cadence
A model that refreshes quarterly or twice a year doesn't match how fast budgets move. Aim for frequent reads you can act on. Look at marginal efficiency, not average returns, since response curves show where a channel starts to saturate. Use what-if scenarios to test moves like shifting $1 million from Meta to YouTube. Haus’ Causal MMM can also recommend which experiments to run next and signal when to rerun one.
For daily decisions, Haus’ Causal Attribution uses experiments to debias attribution, so your read doesn't rest on clicks and views alone. To see how teams operationalize all this, read about how to turn your MMM into a decision engine.
Conclusion
Implementation often gets decided before the model runs. Start with a clear budget decision, build your experiment practice, and structure channels around how you plan and what you've tested. From there, the work is keeping tests fresh and checking that the model agrees with them.
For more on any of these steps, browse the resources page.
Frequently asked questions
What should I decide before I start an MMM?
Decide which budget decision the model needs to inform. Good examples are whether you're overspending on branded search or where extra Q4 budget should go. A model with a clear job often beats a catch-all build, which tends to end in analysis paralysis.
How do I know my team is ready for an MMM?
Three signs point to readiness. You're managing a cross-channel budget, you need visibility into channels you can't geo-segment, and your team acts on experiment results. Are You Ready for An MMM? walks through each one.
Do I need experiments before I build an MMM?
Most teams add an MMM on top of experimentation, not in place of it. You don't need full test coverage to begin, though. Causal MMM uses GeoLift where tests exist and extends estimates through Conversion Lift studies, and testing the channels you can test also improves estimates for the ones you can't.
How granular should my channel structure be?
Build at the level where your team actually moves budgets, and let what you've tested set the ceiling. As a rule of thumb, roll anything under about 2% of your media into a broader category until it's big enough or tested enough to stand alone. Start broad, then add detail as you run more tests.
What should I do if my MMM and my experiments disagree?
Treat it as a signal to investigate. A model that tells you to scale a channel your tests show isn't incremental creates a trust problem. Ben Bolognini, OluKai's VP of Ecommerce, says, "When clear experimentation and directional measurement results disagree, incrementality should win."
Should I build my MMM in-house?
It can look easy at first, since an open-source model plus a small team can produce response curves in weeks. The trouble tends to show up later, as discrepancies across MTA, MMM, and experiments, then prior tuning, then eroding credibility. Around month 12, many teams have to choose between heavy investment in infrastructure and staff or an outside provider.
How often should I refresh the model and rerun tests?
A model that refreshes quarterly or twice a year doesn't match how fast budgets move. There's no fixed cadence for tests, but more frequent channel-level re-runs plus seasonal adjustment can help keep results trustworthy. Sunday, for example, gets a causal read every Monday morning.
Related reading
