Why Your MMM Disagrees with How You Actually Plan

Why your MMM disagrees with how you actually plan

The model should adapt to the way your business actually decides — which goes as far as your experiments allow

At Haus, we often talk about trust in and trust out. The basic idea is simple: if you feed an MMM bad data, you get bad recommendations back. Traditional MMMs are often built on noisy, correlational inputs confounded by multicollinearity, which is one reason so many marketers struggle to act on them. 

It’s also one reason MMMs often go unused. As Haus Solutions Consultant Hannah Perez put it in a recent Haus piece, many enterprise teams end up with an MMM that feels more like an artifact than a decision-making tool. 

But there’s another reason teams struggle to act on MMMs: the model is often speaking a different language than the team using it. Decision-making slows in the translation.

An example: Say your MMM tells you to shift budget from Meta to YouTube. Fine. But what if your team doesn’t actually plan budgets that way? Maybe Meta isn’t one bucket in your planning process. Maybe your team breaks it out into Meta Prospecting and Meta Retargeting. Maybe Google isn’t one line item either, but four: Search Brand, Search Non-Brand, PMAX, and YouTube. If the model reports at a broader level than the team uses to make tradeoffs, the recommendation may be technically interesting but operationally weak.

That’s the granularity problem: Many MMMs are built at a level that does not match how teams actually set budgets, evaluate tradeoffs, or run tests. From the Haus point of view, an MMM becomes more useful when channel definitions reflect real planning decisions and are constrained by the level of causal signal available to calibrate them. Here we’ll dig deeper into why this is important.

The granularity mismatch in MMMs

Granularity is the level at which your model groups marketing activities. Your finance team may think of Pinterest as one line item. Your media team may optimize Meta through prospecting and retargeting. Your growth lead may care most about which budget buckets are actually fluid from month to month. These are all real planning layers. If your MMM doesn’t align with them, it becomes harder to act on what it says.

This is the key question: not how granular can the model get, but at what level does your team actually make decisions? That question matters because more granularity is not always better. In fact, it often makes things worse.

A model that breaks Meta into five sub-channels may look more sophisticated than one that reports a broader Meta bucket. But if your business only really reallocates at two levels, or if only one or two of those sub-buckets have enough credible signal behind them, the extra detail doesn’t make the output more actionable.  It spreads the same evidence thinner; less signal behind each bucket, wider uncertainty, and less you can actually act on.

Haus believes the best solution is to align model structure to planning granularity and test granularity, then increase detail as signal improves.

Your model should reflect how you actually plan

The goal is not to make the model look maximally detailed. The goal is to make it maximally usable.  That's why Causal MMM channel structure has to balance two levels: the planning and budgeting granularity you want, and the tested granularity your evidence can actually support. When they don't line up, the tested level is the binding constraint, it sets how granular you can credibly go, while planning granularity is the target you grow into as experiments accumulate.

Planning and budgeting granularity means the model reflects the buckets where spend is truly fluid. If your team actually makes budget decisions between Meta Prospecting and Meta Retargeting, that may be the right level to structure the model. If your team really plans Google across Search Brand, Search Non-Brand, PMAX, and YouTube, those may be the right buckets too. 

Tested granularity means the model reflects the level where you have credible experimental evidence. This matters just as much. If you want tactic-level recommendations, you need tactic-level signal. If you only have broad, blended evidence, then the model may need to start broader as well. 

This is also where the model earns its keep. 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 come in. A broader starting structure isn't the model giving up detail; instead it's the model being honest about where the evidence is strong and where it's still borrowing signal.

In fact, this is where many MMMs become less helpful than they should be. Teams often ask the model to answer the most detailed version of the question before the evidence exists to support that level of detail. That creates a false sense of specificity.

If your experiments and channels do not line up, your model will not either

Put more bluntly: you can’t ask a blended test to produce unblended truth.

If a brand wants separate readouts for Meta Prospecting and Meta ASC, but the only evidence available is an all-up Meta test, that test may not be enough to support the split. In that case, a broader starting structure is not a compromise. It’s good model design. 

Imagine a brand that budgets Google as Search Brand, Search Non-Brand, PMAX, and YouTube. It budgets Meta as Prospecting and Retargeting. It spends only a sliver on a few smaller paid social channels. In that case, the most useful model probably mirrors those major decision buckets and rolls smaller channels into broader categories until more signal exists. A useful rule of thumb: build channels around distinct, noticeable buckets of spend, and roll anything under ~2% of media into a broader category until it's big enough or tested enough to stand on its own. That may look less granular than a fully exploded model. But it will be much easier to act on. 

This is why it’s important to consider your measurement stack as puzzle pieces that fit together, not each dataset as its own separate puzzle. When you look holistically across how you're measuring, then the data you create maps to the decisions you need to make.

Why this becomes a trust problem

And this is where granularity reconnects back to trust.

 In our experience, when executives say they don't trust MMM, what they often mean is that they don't trust recommendations they can't operationalize.If a model reports at the wrong level, it creates translation overhead. Teams spend time interpreting the output before they can even debate whether they believe it. That’s how an MMM becomes a reporting artifact instead of a decision engine.

Marketers don’t reject MMM because they hate analytics. They reject models that don’t map to the real mechanics of planning.

Start where decisions are made

So what should teams do instead?

Start where decisions are actually made. Build the model at the level where budgets move. Calibrate it with causal signal at that level. Then, as the evidence improves, increase granularity over time. 

That’s a healthier path than forcing the most granular version of reality on day one and hoping the model can support it.

The most useful MMM is not the one with the most rows in the output. It’s the one that helps a team make a real decision with confidence.

An MMM should not force your planning process to adapt to the model. The model should adapt to the way your business actually decides - as far as your experiments allow, and then keep closing that gap test by test.

‍