Marketing mix modeling spent decades as an enterprise ritual: a consultancy, a six-figure engagement, a deck that arrived twice a year and mostly confirmed what the CMO already believed.
Then privacy changes broke user-level tracking, attribution got murkier every quarter, and MMM came back into fashion, because it’s the one measurement method that doesn’t care about cookies, consent banners, or walled gardens at all. Meta open-sourced Robyn. Google built Meridian. A wave of vendors put the whole thing behind a SaaS login.
So now MMM is accessible. What hasn’t changed is the content about it, which still splits into two piles: vendor pitches that promise a truth machine, and academic material that assumes you know what a Bayesian prior is and why you should care.
This is the third leg of a series: the measurement stack covered the system, the incrementality playbook covered experiments, and this covers the modeling layer: what it actually does, when it’s worth your time, and how to sanity-check one without a statistics degree.
What MMM Actually Does (In One Paragraph)
MMM works top-down. It takes your historical weekly data (spend by channel, revenue or conversions, plus context like seasonality, promotions, and pricing) and estimates how much each input contributed to the outcome. No user tracking, no pixels, no attribution windows. Just the statistical question: when spend on this channel moved, how did revenue move, holding everything else as constant as the math allows?
That framing explains both its superpower and its weakness. The superpower: it sees everything, including the stuff attribution is blind to (TV, podcasts, brand search absorbing other channels’ demand, the halo from that PR hit). The weakness: it’s correlation-based inference on messy observational data, which means it can be confidently wrong, and it will not warn you when it is.
When It’s Worth It (And When It’s Not)
MMM has a real cost even when the software is free: data assembly, setup, and the ongoing work of interpreting it. The honest thresholds:
MMM starts earning its keep when:
- You spend enough for the statistics to work. As a rough floor, think $1M+ per year across three or more channels. Below that, the model doesn’t have enough variation to learn from.
- A meaningful share of your spend is hard to measure directly: TV, audio, out-of-home, sponsorships, top-of-funnel video.
- You’re making quarterly or annual allocation decisions across channels, which is the decision MMM is actually built for.
- You have at least two years of weekly history (or can assemble it).
Skip it (for now) when:
- You’re spending on one or two channels. A blended CAC view plus periodic incrementality tests will answer your questions faster and more credibly.
- Your spend levels barely change. MMM learns from variation; if every channel gets the same budget every week, there’s nothing to learn from.
- You want campaign-level or creative-level answers. MMM operates at the channel-and-quarter altitude. It will never tell you which ad worked.
The failure mode I see most isn’t teams skipping MMM when they need it. It’s teams adopting it too early because it sounds rigorous, then treating noisy output as strategy.
The Tooling Landscape, Briefly
- Google Meridian and Meta Robyn: open-source, free, and genuinely good. The catch is that “free” means someone technical (a analytically-strong marketer, a data-adjacent engineer, or a contractor) spends real days on data prep and setup. Meridian is the more opinionated and currently better-supported path; Robyn is more configurable.
- SaaS MMM vendors: they compress setup time and give you a maintained dashboard. What you’re paying for is convenience, not better math; the model quality is bounded by your data either way. Judge them on transparency: if a vendor can’t or won’t show confidence intervals, explain their assumptions, or let you export the underlying model, walk.
- A spreadsheet regression you build yourself: educational, and legitimately useful for building intuition, but don’t route budget decisions through it. The hard parts of MMM (adstock, saturation, seasonality, avoiding spurious fits) are exactly the parts a naive regression skips.
For most teams reading this, the pragmatic answer is Meridian with contractor help for setup, or a transparent vendor if nobody internal wants to own it.
The Real Bottleneck Is Your Data
Every MMM disappointment I’ve seen traces back to inputs, not algorithms. What the model needs:
- Two-plus years of weekly spend by channel, from your actual invoices and platforms, not from memory. Agency-managed channels and one-off sponsorships are where the holes usually are.
- One outcome variable you trust: revenue or new customers from your backend or CRM, the same source-of-truth discipline as everywhere else in the stack.
- The context that explains the outcome: promotions and their dates, price changes, seasonality, stockouts or capacity limits, major PR moments, and anything else that moved revenue for non-media reasons. If it’s not in the model, the model will hand its credit to whichever channel happened to be spending at the time.
Assembling this is unglamorous and takes longer than the modeling. It’s also reusable: the same clean dataset feeds forecasting, planning, and the CAC analysis you should be doing anyway.
How to Tell If Your MMM Is Lying: Five Checks
You don’t need to audit the math. You need to interrogate the outputs like an operator.
- The smell test. If the model says a channel’s true ROI is triple what any other evidence suggests, the burden of proof is on the model. MMMs love to over-credit channels that correlate with demand (brand search is the classic), and under-credit slow-building ones.
- Confidence intervals, always. A channel ROI of 2.4 with a range of 1.9 to 2.8 is actionable. A ROI of 2.4 with a range of 0.3 to 5.1 is the model saying “I don’t know” in a confident voice. If your tool doesn’t surface ranges, that’s disqualifying.
- Do the saturation curves make sense? MMM estimates diminishing returns per channel. If the curve claims you could triple Meta spend at constant efficiency, or that you saturated search at a spend level you’ve exceeded happily for a year, the model has fit noise.
- Holdout accuracy. Any competent setup tests the model against a period it wasn’t trained on. Ask for that number. If the model can’t roughly predict last quarter, it has no business allocating next quarter.
- The experiment cross-check. This is the big one, covered next.
Triangulation: Experiments Keep the Model Honest
The strongest measurement setups don’t pick a method; they use three that check each other, each at its own altitude:
- Attribution (daily/weekly): fast, biased, fine for spotting changes and steering ops.
- Incrementality tests (per big decision): slow, narrow, and causal. The anchor for truth.
- MMM (quarterly/annual): the wide view that includes everything, calibrated against the other two.
The practical loop: when your MMM makes a strong claim about a channel, test it. Run a geo holdout or scale-back experiment on that channel and compare the measured lift to the model’s estimate. When they roughly agree, your confidence in the whole system compounds. When they disagree, believe the experiment, and feed the result back into the model (Meridian and most serious vendors support experiment-based calibration directly, and it’s the single highest-leverage configuration step there is).
A model that has never been checked against an experiment is an opinion with error bars. A model calibrated against two or three real tests per year is a planning tool you can defend in a budget meeting.
A Crawl, Walk, Run Path
Crawl (this quarter): Assemble the dataset. Two years of weekly spend by channel, a trusted outcome variable, and a written log of promos, price changes, and anomalies. This is valuable even if you never build the model.
Walk (next quarter): Stand up a first model (Meridian, or a transparent vendor). Run the five checks. Use the output for one decision class only: directional reallocation questions at the quarterly level, in combination with what your experiments and blended metrics already say. Don’t let it near weekly budgets.
Run (this year): Calibrate against at least two incrementality tests. Refresh the model quarterly. Let it own the annual planning conversation: channel mix, saturation-aware budget scenarios, and the “what would happen if we cut TV” questions that attribution was never able to answer.
Closing Thought
MMM’s return isn’t a fad; it’s the logical response to a world where user-level tracking keeps degrading and top-down inference is what’s left. The accessibility is real: what took a consultancy and six figures in 2018 is now an open-source library and a motivated analyst.
But accessibility cuts both ways. It’s never been easier to build a model, which means it’s never been easier to build a bad one and put a number with two decimal places on a slide. The discipline that separates useful MMM from expensive theater isn’t statistical sophistication. It’s operator skepticism: clean inputs, confidence intervals, sanity checks, and experiments that keep the model honest. Same as every other layer of the stack.