MMM Is a Rearview Mirror, and Four Other Things People Get Wrong About It

Most objections to Marketing Mix Modeling come from a version of it that is ten years out of date. Here are the five that come up most, and what is actually true now.

Myth 1: "MMM only tells you what already happened"

This is the most common one, and it describes how MMM used to be deployed, not what it does. A model built on your historical spend and revenue does explain the past — but the same relationships it learns are what a forecast runs on. Once the model knows how each channel has responded to spend, it can project what next quarter looks like under a budget you have not spent yet, including budgets you have never tried. That is a forecast, not a rearview mirror. The "backward-looking" complaint is really a complaint about models that were refreshed once a year and never looked at again. A model that refreshes as new data lands is forward-looking by construction, because every refresh updates what it expects next.

Myth 2: "You need years of data before it's useful"

Less than most teams assume. A model can run on as little as 52 weeks of weekly spend and revenue data, because that is enough to see a full seasonal cycle once. What matters more than the number of weeks is whether your spend has actually moved during that window — a channel that has sat at the same budget for a year gives a model very little to learn from, while a channel that has been tested at three different budget levels teaches it a lot, even in less time. Two to three years of history produces the most stable results, but waiting for that is usually the wrong trade: a model built on one good year, recalibrated against a live incrementality test, outperforms no model at all by a wide margin. We cover what "enough" data actually looks like in the companion piece on data history.

Myth 3: "It can't see creative quality"

True, and worth saying plainly. MMM measures what a channel did at the spend level you gave it — it does not know whether the ad creative running in that channel was good or bad. What it can tell you is whether the channel's return changed when the creative changed, if you treat a creative refresh as a shift in the data the same way you would treat a budget shift. MMM and creative testing answer different questions: one tells you which channel to fund, the other tells you what to put in it. Teams that expect MMM to grade their ads are asking a budget-allocation tool to do a creative team's job.

Myth 4: "MMM replaces attribution, or attribution replaces MMM"

Neither replaces the other, because they are not competing for the same decision. MMM operates at the level of quarterly and monthly budget allocation across channels — it is built to answer "should more money go to YouTube or paid search this quarter." Attribution operates at the level of which touchpoint, campaign, or ad set gets credit for a specific conversion — it is built to answer "which creative in paid search is actually converting." A team that only runs MMM has no way to optimize within a channel week to week. A team that only runs attribution has no way to catch a channel that is reporting a strong ROAS on paper while barely driving anything incremental. Run both, and let each one check the other: when attribution says a channel is converting well but MMM says its contribution has flattened, that gap is where over-credited channels usually hide.

Myth 5: "It's a black box you have to trust blindly"

This one is fair criticism of older tools, not of the category. The reasonable version of this complaint is: a single ROAS number with no indication of how confident the model is in it is not something a finance team should sign off on without question. The fix is not to abandon modeling — it is to demand a model that reports a range instead of one number, so you can see whether "Meta ROAS is 3.2x" is a tight, well-supported estimate or a wide guess that needs another data point before anyone acts on it. A model that also checks its own stability before showing a recommendation, and that updates its range every time a new incrementality test comes in, is answerable in a way a single static number never was.

The actual question to ask

None of these five objections are reasons to skip measurement — they are reasons to ask better questions of whatever measurement you use. Does it forecast as well as it explains? How much spend movement does your data actually have, not just how many weeks? What is it silent on, and have you paired it with something that covers that gap? Does it tell you which channel to fund, or which ad within a channel to run, or both? And does every number come with a range, or just a number? Those questions apply whether you build the model yourself or run it on a platform — they are just easier to answer when the platform shows its range on every estimate instead of asking you to trust a single line.