A five million pound brand with four clean years of history is better positioned for this than a forty million pound brand that replatformed last spring.
Marketing mix modeling has a reputation as an enterprise technique, and until recently that reputation was earned: it meant six-figure consulting engagements and proprietary software. That changed when the largest advertising platforms open-sourced credible implementations, and the gate moved.
Where the gate moved to is worth understanding precisely, because the common framing of the question is wrong. Brands ask whether they are big enough. Size is a proxy for the thing that actually matters, and it is a poor one.
Why MMM Came Back
The technique dates to the 1960s. Its return has nothing to do with the mathematics improving and everything to do with the alternative degrading.
Multi-touch attribution depends on identifying individual users across sessions and devices. Privacy controls, tracking prevention and consent requirements have made that identification unreliable, with reported figures putting attribution identity coverage falling from above ninety percent to somewhere around thirty to sixty percent.
Marketing mix modeling. A statistical technique that estimates the contribution of each marketing activity to a business outcome using aggregated time-series data, typically weekly spend per channel against revenue, alongside control variables such as seasonality, pricing and promotions. Because it operates entirely on aggregate figures and requires no user-level data, it is structurally unaffected by cookie deprecation, consent rates, or platform tracking restrictions, which is the property that has driven its return.
That last property is the whole argument. MMM does not need to know who anyone is. It observes that spend went up in one channel during certain weeks and that revenue behaved a particular way, and it estimates a relationship. Nothing in the privacy landscape touches that.
The second driver is commercial. Finance teams are asking harder questions about which channels work, and platform-reported return on ad spend is a poor answer because every platform claims the same conversions. MMM produces one cross-channel view that does not come from any party with an interest in the result.
Observations, Not Revenue
The reframe that makes this question answerable.
A statistical model needs enough data points to distinguish signal from noise. The number of data points you have is a function of how long you have been collecting and at what granularity, not of how much you sell. Revenue enters only indirectly, in that larger brands tend to have longer histories and more channels.
Two brands with identical revenue can be years apart in readiness. The one that never changed its channel mix and kept clean weekly records can model. The one that replatformed and switched agencies cannot, whatever it turns over.
This also explains why the advice you find varies so much. Sources recommending two years, three years, or eighty to a hundred observations are all describing the same underlying requirement from different angles, and none of them can tell you your answer because your answer depends on your channel count.
The practical consequence is encouraging for smaller brands. A focused business running four channels consistently for three years is in better shape than a larger one running twelve channels for eighteen months, because the first has more observations per variable to work with.
Calculate Your Own Requirement
A rule of thumb that takes a minute to apply and answers the question better than any revenue threshold.
The commonly cited version requires at least ten observations per independent variable. Your independent variables are your channels plus your control variables, and control variables include things like seasonality, pricing, promotions and any major events you need the model to account for.
Note what the arithmetic implies about channel proliferation. Every additional channel you add raises the data requirement, which means a brand spreading spend thinly across many channels needs substantially more history than one concentrating on a few.
Set against that, the broader guidance recommends around three years of weekly data, roughly 156 observations, on the grounds that this covers two full seasonal cycles and several promotional waves. That is a different argument from the observations-per-variable one: it is about seeing your seasonality repeat rather than about statistical power.
Both matter. Clear the observations-per-variable bar to have a stable model, and cover two seasonal cycles to have a model that understands your year.
Three Ways To Get Enough Data
Most coverage mentions only the first of these, which is why smaller brands conclude they must wait years. Two others exist.
Accumulate calendar time. Two years of weekly data as a minimum, three years recommended. Monthly data requires three to four years and is worse in every respect.
Model across regions or segments rather than nationally. Cross-sections multiply effective observations and compress the calendar requirement substantially.
For direct response businesses, the authors of Robyn note that daily observations over shorter periods can suffice to estimate an informative model.
A direct-to-consumer brand with daily data across several regions can reach viable observation counts far sooner than the standard three-year advice suggests.
On pooling, one source gives a concrete worked figure: eighteen months of weekly data across six regions carries 468 region-weeks of evidence, described as statistically comparable to a three-year national dataset. That is a substantial shortcut for any brand whose sales vary geographically in ways it can observe.
On daily data, the point comes from the Robyn authors themselves. Their paper describing the framework notes that while a minimum of two years of weekly observations is usually recommended, in direct response business models such as direct-to-consumer, daily data observed for shorter time periods can suffice to estimate an informative model.
That qualifier is directly relevant here and rarely quoted. Ecommerce brands are direct response businesses. The standard advice was written with the general case in mind, and the tool's own authors say the general case may not bind for you.
What The Model Needs From You
Data volume is necessary and not sufficient. The inputs have to be clean, aligned and complete.
- Spend per channel per period. Actual spend, consistently defined, with no gaps and no periods where a channel's reporting changed definition.
- An output variable. Revenue or conversions, from your own systems rather than from any platform's attributed figures.
- Control variables. Seasonality, price changes, promotional periods, and major events such as launches or outages.
- Time alignment. Every series on the same calendar, with the same period boundaries. Weeks that start on different days across sources produce nonsense.
- Impressions or clicks alongside spend where available, since spend alone conflates media cost changes with media volume changes.
The last point is worth dwelling on. If your cost per thousand impressions rose forty percent while your spend stayed flat, you bought considerably less media, and a model seeing only spend will interpret that period incorrectly. Including a volume measure lets the model separate price from quantity.
Not model complexity. Data assembly. Reconstructing three years of consistently defined weekly spend across every channel, when agencies changed, accounts were restructured, and platform reporting definitions moved, is genuinely difficult work and it is where most projects stall. Assess whether you can build the dataset before evaluating any modeling tool.
The Three Open-Source Options
Three credible, free, production-grade frameworks exist, which is what changed the economics.
| Framework | Origin And Language | Notes |
|---|---|---|
| Robyn | Meta, R | Uses gradient-free optimization for hyperparameter tuning. Strong community. Reportedly not designed for weekly updates or real-time decisions. |
| Meridian | Google, Python | Bayesian. Launched globally in January 2025, with a no-code scenario planning interface reported in February 2026. Deeper integration with Google's ecosystem. |
| PyMC-Marketing | PyMC Labs, Python | Bayesian, built on a general probabilistic programming library. Most flexible and least opinionated. |
Both platform-backed frameworks analyze all channels regardless of platform, so neither locks you into its parent's ecosystem. That said, each was built by a company with a channel it would like to look good, which is a reason to cross-check rather than a reason to avoid them.
Sophisticated teams reportedly run Robyn and Meridian together, using each as a check on the other. That is sensible practice and it doubles the engineering burden, so it belongs to organizations with data science capacity rather than to a first attempt.
For a first model, choose on the language your team already uses. A team fluent in Python should not learn R to run Robyn, and a team with neither should be honest that this is a hiring or contracting decision rather than a tooling one.
What The Output Actually Is
Worth being precise, because expectations set by attribution dashboards do not transfer.
MMM produces estimated contributions by channel over the modeled period, along with response curves showing how contribution changes as spend changes, and typically a recommended allocation. Those estimates come with uncertainty intervals, and the intervals are part of the answer rather than a technicality.
What it does not produce is anything about individual customers. There is no journey, no user path, no ability to look up why a particular order happened. It answers portfolio questions only.
The most valuable output is usually not the point estimate of any channel's contribution but the shape of the response curves, because those tell you about diminishing returns. Knowing that a channel's marginal return falls sharply above a certain weekly spend is a directly actionable finding, and it is not available from attribution at all.
Treat the uncertainty seriously. A model estimating a channel contributed somewhere between five and thirty percent of revenue is telling you it does not know, and reporting the midpoint as though it were a measurement misrepresents your own analysis.
The Cost Nobody Quotes
Free frameworks changed the licensing line and left the rest of the cost structure intact.
One source puts it well: the open-source route is free on the licensing line and expensive on the engineering line. The realistic cost sits in data pipelines, infrastructure, validation discipline, and the specialist expertise to configure priors and transformations responsibly.
That last item deserves attention because it is the least visible. Bayesian models require choices about prior distributions and about how spend is transformed to reflect diminishing returns and delayed effects. Those choices materially affect the answer, they are not obvious, and getting them wrong produces a model that looks fine and is wrong.
Consultancy: buys advisory depth and methodological rigour at high fixed cost, usually on a quarterly cadence, with the methodology staying vendor-side. You rent the capability rather than building it. Open source in-house: free to license, expensive in engineering and expertise, and you own it. Commercial software: sits between the two, competing now on usability and decision support rather than on model quality, since the open-source floor rose.
The ongoing burden is also underestimated. Models need re-running as campaigns and channels evolve, and the frameworks are not built for weekly refreshes, so this is a quarterly discipline rather than a dashboard. A brand expecting a live view of channel contribution has misunderstood what it is buying.
The Ecom Profit Box
Our collection of ecommerce growth resources, including the margin frameworks any allocation decision needs.
Get It FreeDo You Have The Data?
We do not build models. We can help you work out whether the dataset exists before you commit to a project.
Book A CallReading The Benchmark Claims
Numbers circulate about what MMM delivers, and they need the usual treatment.
One figure quoted in industry coverage attributes to a vendor benchmarks report the claim that optimized media planning can shift marginal return on investment from below one dollar to $5.84 when decisions are grounded in planning data rather than platform-reported return.
Read the construction of that claim. It describes what optimized planning can achieve, sourced from a company selling planning software, with no methodology available and no indication of how typical the outcome is. It is a best case presented as a benchmark.
The plausible core is that reallocating budget away from channels with exhausted marginal returns toward channels with headroom improves efficiency. That is the mechanism MMM supports and it is genuinely valuable. The magnitude in any specific case depends entirely on how misallocated you were to begin with.
Which suggests the honest expectation. If your current allocation was set by platform-reported return, MMM will probably tell you that some channels are less productive than they appear, and acting on that will improve efficiency by an amount proportional to how wrong you were. Brands already allocating thoughtfully should expect smaller gains.
How It Fails
Specific failure modes, most of which produce a model that looks credible.
- Correlation presented as contribution. If a channel's spend always rises in December alongside seasonal demand, a model without adequate seasonality controls will credit the channel for the season.
- Insufficient variation. A channel whose spend never changes offers the model nothing to learn from. Constant spend means an unidentifiable effect.
- Collinear channels. Two channels that always move together cannot be separated. The model will apportion between them arbitrarily and report both with confidence.
- Incomplete history. Missing or inconsistently defined periods produce unreliable channel contribution estimates, and the model will not tell you that is why.
- Over-fitting to the past. A model explaining historical revenue perfectly may predict nothing, which is why out-of-sample validation is not optional.
- Treating output as measurement. These are estimates with uncertainty, not observations.
The collinearity problem is the one that most affects smaller brands specifically. If you have always increased all spend together during promotions and decreased it together afterwards, your channels are correlated with each other and the model cannot distinguish their effects, regardless of how much history you have.
The fix is deliberate variation. Changing channel budgets independently of one another, even modestly, creates the variation a model needs. That has a cost and it is an investment in being able to measure at all.
Where It Sits In A Stack
MMM is not a replacement for the other measurement methods, and the strongest position combines them.
Attribution traces individual digital journeys and optimizes campaign-level decisions within a channel. Precise on what it can see, blind to offline and to anything without an identifier.
MMM measures the incremental contribution of every channel including offline, at the budget level. It answers allocation questions and cannot answer anything about individuals.
Experiments establish causation directly by withholding exposure from a comparable group. The only method that observes a counterfactual rather than inferring one.
The relationship worth understanding is that experiments can calibrate a model. A holdout result gives you a measured effect for one channel, which can be used to anchor the model's estimate for that channel and improve the whole. Frameworks in this space increasingly include facilities for exactly that calibration, and it is the most valuable connection between the two methods.
Whatever the model says, judge the resulting allocation against contribution margin rather than revenue, since a channel driving revenue at poor margin is not a channel to fund. Our contribution margin playbook covers building that view and the true cost of a sale covers what belongs in it.
Who Should Not Do This
The honest exclusions, and they are not primarily about size.
Brands running one or two channels. If nearly all spend is in one place, you do not have an allocation problem. Your question is how much to spend rather than where, and an incrementality test answers that better and cheaper.
Brands without the history or the cross-sections. If you cannot clear the observations bar through calendar time, pooling, or daily data, a model built anyway will produce confident numbers with no support underneath them.
Brands whose channels always move together. Without independent variation, the model cannot separate effects no matter how much data you supply.
Brands that will not change allocation. If budgets are set by other constraints, the model is expensive information you cannot act on.
Brands without data capacity. The blocker is assembling clean, aligned, complete series across every channel. If nobody owns that, the project stalls there regardless of which framework you chose.
The Practical Sequence
Count your channels and controls, multiply by ten, and compare against your available observations counting weeks, or region-weeks if you can pool, or days if you are direct response. That tells you whether this is available to you now, later, or not at all. Then assess honestly whether you could assemble three years of consistent weekly spend data if asked, because that is the step that actually stops most projects.
If both answers are yes, the frameworks are free and the barrier is engineering rather than licensing. If the first is no, run incrementality tests instead and revisit when the history exists. If the second is no, fixing your data foundation is worth doing regardless of whether you ever build a model, because everything else in measurement depends on it too.
What To Remember
- The constraint is observations, not revenue. A focused brand with a long clean history is better positioned than a larger one that replatformed recently or runs many channels briefly.
- Calculate your own requirement: at least ten observations per independent variable, where variables are channels plus controls. Eight channels and four controls implies roughly 120 weeks.
- Pooling across regions multiplies observations. Eighteen months of weekly data across six regions is 468 region-weeks, reportedly comparable to three years of national data.
- Direct response businesses may use daily data. The authors of Robyn note that in models such as direct-to-consumer, daily observations over shorter periods can suffice.
- Open source is free on licensing and expensive on engineering. The cost is pipelines, validation discipline, and the expertise to set priors and transformations responsibly.
- MMM needs no user-level data, which is why privacy restrictions and consent rates do not degrade it, and why it returned as attribution identity coverage fell.
- Channels that always move together cannot be separated, so deliberate independent variation in budgets is a prerequisite for measuring anything.
Where This Came From
- The Robyn framework paper on arXiv, authored by the framework's developers, which describes the recommendation of a minimum of two years of weekly observations and notes that in direct response business models such as direct-to-consumer, daily data observed over shorter time periods can suffice to estimate an informative model.
- Robyn repository, Google Meridian documentation, and PyMC-Marketing documentation.
- Analytics industry reporting on data requirements, including 80 to 100 observations for a stable model, roughly two years of weekly history as a minimum and three years as the standard recommendation covering two seasonal cycles, monthly data requiring three to four years, and the rule of thumb of at least ten observations per independent variable.
- Consultancy reporting on pooling across cross-sections, including the worked figure that eighteen months of weekly data across six regions carries 468 region-weeks of evidence described as statistically comparable to a three-year national dataset, and the characterisation of the open-source route as free on the licensing line and expensive on the engineering line.
- Industry reporting on Google launching Meridian globally in January 2025 and adding a no-code scenario planning interface in February 2026, on Robyn not being designed for weekly updates or real-time budget decisions, and on multi-touch attribution identity coverage falling from above 90 percent to roughly 30 to 60 percent.
- A vendor benchmarks report cited in industry coverage claiming optimized media planning can shift marginal return on investment from below one dollar to $5.84. This originates from a company selling planning software, publishes no accessible methodology, and describes a best case rather than a typical outcome; section 09 discusses it on that basis.