Marketing mix modeling has been having a good few years, and the pitch is appealing. It measures every channel without a single cookie, it survives privacy rules that broke click tracking, and open-source frameworks mean you no longer pay a consultancy six figures to get started. Marketers at smaller brands hear all of that and ask a reasonable next question: should we be doing this too.
The honest answer is often no, and not because of cost. MMM is a statistical model. A model needs enough signal to learn from, and a smaller brand frequently cannot supply it. Buy MMM before you can feed it properly and you do not get a smaller, rougher version of the enterprise result. You get confident-looking numbers built on noise, which is worse than no numbers at all. This piece is about telling which side of that line you are on, before you spend anything.
Why MMM has a real floor
Start with what the model is doing. MMM relates a single outcome, usually weekly sales or revenue, to everything that might explain its movement: spend on each channel, plus non-marketing factors like price, promotions, seasonality and competitor activity. It is usually a regression on aggregate data. From the pattern of "spend moved like this, sales moved like that," it works backward to an estimate of what each channel contributed.
That backward step is only possible if the data holds enough information to support it. The model has to do three hard things at once: separate each channel's effect from random week-to-week noise, separate each channel's effect from every other channel's, and separate all of marketing from the baseline sales the brand would have made anyway. Each task consumes signal. A big advertiser with many channels, years of history and spend that moves around has signal to spare. A small brand with three channels, eighteen months of data and a steady budget does not, and the model cannot manufacture what is not there. It will still return an answer, just not one that means much.
This is the part the marketing of MMM tends to skip. The technique has a minimum scale built into its mathematics, and no software removes it.
The budget rule worth knowing
The most quoted threshold is a cost ratio. Measurement firm Circana puts it directly: a brand's marketing budget should be at least fifty times the cost of the MMM project for the investment to make sense. The logic is plain. If MMM helps you reallocate budget, the value it creates is a fraction of the budget it steers, so a fixed model cost against a small budget eats the gains.
A practitioner write-up on MMM pricing makes the same point with numbers. If you spend 100 million dollars a year on marketing, a 100,000 dollar model is 0.1 percent of spend, trivially worth it if it sharpens allocation even slightly. For a brand spending 1 million a year, that same project is a tenth of the annual budget, and no model can return that.
Open-source frameworks change this arithmetic less than people assume, and we will come back to why. For now, treat the fifty-times rule as a first filter. If your budget is not comfortably above fifty times what a credible MMM effort would cost, in money and staff time, the model is unlikely to pay for itself yet.
There is a simpler version of the same test. One guide for smaller companies suggests that if you spend under roughly 100,000 dollars a year across fewer than four channels, you should fix tracking and attribution first and leave MMM for later. Not a hard line, but a fair gut check.
The data rule, and an honest disagreement about it
Budget is the cost side. Data is the can-it-even-work side, and here the model is unforgiving.
The standard requirement is two to three years of history at a weekly grain. The documentation for Meta's Robyn, the framework that did more than any other to popularise open-source MMM, states that a model needs a minimum of two years of historical weekly data, and if only monthly data exists you should collect four to five years instead. Consultancy Artefact gives the same two-year weekly guidance. The reason is seasonality. Sales for most brands rise and fall on an annual cycle, and the model has to see that cycle repeat a few times to tell a real channel effect apart from a December that was always going to be busy. Two years of weekly data is only about 104 data points, a small sample for a model untangling several channels plus price, promotions and seasonality.
That leads to a rule worth carrying around. Robyn's guide recommends ten observations for every independent variable in the model as the target ratio. Artefact frames it as needing at least three times more data points than parameters to avoid overfitting. Either way, the count of things you are trying to measure is tied to the history you need. More channels do not just give you more to learn, they raise the bar for how much data you must already have.
In fairness, the requirement is contested. A wave of lower-cost MMM vendors argues the two-year rule is too conservative. One says six months of data is enough to start, with anything below that treated as exploratory. Another recommends twelve months of daily data with a target metric in the hundreds every week. They have a real point: years alone do not make data adequate. What the model needs is variation, and a year where spend moved around can beat three years where it sat still. True. It is also the argument a vendor makes when a longer requirement would shrink the addressable market. The cautious read, the one that protects you from a bad model, is that two-plus years of weekly history is the safe target, and anything shorter is exploratory rather than something you bet a budget on.
Why two or three channels is not enough
Budget and history can both clear the bar and the model can still fail, because of how few channels smaller brands tend to run.
Picture a brand on three channels: Meta, Google and email. The trouble is rarely the count itself. It is that small teams move all three together. The quarterly budget goes up, so spend on everything goes up. The slow season comes, everything comes down. To the model, three channels that always rise and fall in step look like one channel wearing three labels. It cannot tell which one moved the needle, because the revenue can be explained by any of them.
This is multicollinearity, the failure mode that quietly wrecks small-brand MMM. When inputs move together, the regression cannot split the credit cleanly. The coefficients turn unstable: one channel can inflate, another can collapse toward zero, and small changes in the data swing the results around. Practitioners flag trouble when channel correlations climb above about 0.7, or a variance inflation factor above 5. The model still prints an answer, just not one you should reallocate budget on.
It can be reduced, and the fix is in your hands rather than the software's. You break the lockstep on purpose: pulse one channel up one week and a similar one the next, or go dark on one suspected-collinear channel at a time so the model finally sees them move independently. That works, but it means running your media in a way that serves the model, which a brand with three channels and a tight budget may not want or be able to do.
Spend that never moves is invisible
There is a blunter version of the same problem, and it catches people out.
A channel whose spend never changes teaches the model nothing. MMM learns by watching spend go up and down and seeing what sales do in response. Hold a channel flat and there is no up, no down, no response to observe. Robyn's documentation gives the example plainly: if TV activity stays constant while sales vary, the model struggles to work out how TV affected anything. The effect does not come back as a useful zero. The channel's steady contribution gets folded into the baseline, the sales the model assumes you would have made with no marketing at all. Your always-on channel can be doing real work and the model will hand its credit to thin air.
This is why three years of flat spend on every channel can produce worse attribution than nine months where spend genuinely varied. It also reframes what counts as enough data. The real question is how many weeks each channel was both active and changing. A channel that ran at exactly the same number every week for two years contributes no usable signal, however long the file is.
What smaller brands should do instead
If MMM is not ready for you, the answer is not to fly blind. Use the methods that work at your scale and cost far less, and grow into MMM when the scale arrives.
Start with experiments, because they give you something MMM on its own cannot: causal proof. The simplest is a holdout. Switch a channel off, or hold back a slice of your audience from seeing it, and compare. The gap is the channel's real incremental contribution, measured rather than modeled. Geo tests apply the same idea across markets, running a channel in some regions and not others. None of this needs years of history or an eight-figure budget, and a small brand can run a clean test in weeks. For how lift and control groups work, see our piece on incrementality testing without the jargon.
Next, use the platform tools you already pay for. Meta, Google and TikTok all offer conversion lift and geo lift tests inside their own ad products. Read them with healthy skepticism, since a platform grading its own work has an obvious incentive, but as a low-cost read on whether a channel pulls its weight they are a reasonable place to begin. A growing set of incrementality testing tools aimed at smaller brands, several built for Shopify, has made geo testing approachable without a five-figure annual contract.
None of this is a downgrade from MMM. Experiments and platform tests are the right tool for your current scale, and they are also the calibration MMM will need later. Even measurement vendors describe the methods as a layered system: experiments provide causal truth, and MMM, once you are big enough, sets strategic allocation. Running tests now builds the track record you will use when MMM finally earns its place.
The open-source caveat
One more thing, because it is the most common reason a smaller brand talks itself into MMM too early. Open-source frameworks are free, and free is a real barrier removed. The actively maintained ones today are Google Meridian and PyMC-Marketing, with Meta's Robyn the influential predecessor that Meta has since stopped developing. The mistake is assuming free software lowers the requirements for everything else. It does not.
The frameworks lower the licensing cost and nothing else. They still expect clean, structured input, which means building and maintaining a data pipeline. They still need someone who can specify a Bayesian model, read its diagnostics and know when it is misbehaving. One analysis notes that a first validated model commonly takes six to sixteen weeks from kickoff to production, with ongoing work after. The software is free. The data engineering, the expertise and the time are not. And none of it touches the data and scale floor. A free framework run on too small a budget and too thin a history produces the same unreliable model as an expensive one, just without the invoice.
A short self-assessment
You can decide most of this yourself. Before considering MMM, answer five questions honestly.
- Is your annual marketing budget comfortably more than fifty times what a credible MMM effort would cost you, in money and in staff time?
- Do you have at least two years of history, ideally at a weekly grain?
- Do you run enough channels, roughly five or more, for the model to have something to compare?
- Has spend on those channels actually varied over that period, rather than holding flat?
- Can you, or will you, run incrementality tests to calibrate the model and check it against reality?
Five clear yeses, and MMM is worth a serious look. Two or three noes, and your budget is better spent on experiments, on platform lift tests and on the tracking foundations that make every later measurement better. That is not settling. MMM rewards scale, and the honest move at smaller scale is to build toward it rather than buy a model that cannot yet see straight. For the wider story of why this old technique came back, see our piece on the MMM renaissance.
Council summary
This post argues that the real barrier to marketing mix modeling is not its price but its appetite for data, and that buying MMM before you can feed it yields confident numbers built on noise, which is worse than no numbers. It hands the reader two concrete tests: the cost side, where a marketing budget should sit comfortably above fifty times the project cost, and the can-it-work side, where roughly two years of weekly history, five or more channels, and spend that genuinely moved are what a regression needs to tell one channel from another. The takeaway for a decision-maker at a smaller brand is to run incrementality and platform lift tests now, treat them as the right tool for the current scale rather than a downgrade, and grow into MMM once the budget, the channels and the variation are all there. The honesty about vendor incentives holds throughout: it flags that shorter data requirements are the argument a vendor makes when a longer one would shrink its market, and that free frameworks remove the licensing fee and nothing else.
Comments