Marketing Mix Modeling: What It Costs Before It Tells You Anything

You are in a budget meeting, and the slide orders channels by return. Yours is at the bottom. It is not there because it performed worst. It is there because nobody in the room can prove what it did, and a channel you cannot prove loses to a channel you can. So the budget moves, and the year you spent learning that channel becomes money somebody else gets to spend.

You have signed off on that trade yourself. I have too. It gets filed under data quality and never gets fixed, because the fix costs more than the channel did.

The technique built for this problem, marketing mix modeling, carries a price and a data requirement that rarely reach the person approving the budget, and neither tends to be on the slide when the decision gets made.

What decides whether it is worth buying? Not the number the vendor quotes on the call, and not the years-of-history figure a guide will quote you.

Why Marketing Mix Modeling Is Expensive Before It Is Useful

One clarification before the costs. Marketing mix modeling here is the statistical method that separates a channel's contribution from everything else moving at the same time, not the four Ps of product, price, place and promotion.

The published price is the visible half of the cost. The other half lands before the model produces anything you can act on, in three places.

Data assembly is the first. Your spend, your sales and your external variables live in different systems at different grains. Someone has to reconcile them into one weekly table, decide what to do with the quarter your CRM changed, and document it. Odins, a vendor guide, puts this at two to six weeks to assemble the history, then hours to days every month after that, for as long as the model runs. In the business cases I have seen, that monthly work is the line that goes missing.

Talent is the second. Open source frameworks like Meridian, Robyn and PyMC carry no license fee, which makes them look free until you cost the person who runs them. Cassandra's guide, published by a vendor that sells managed MMM, puts the in-house equivalent at $100,000 to $200,000 or more in data science headcount. That is the comparison a buyer has to make before the open source route can be called free.

Then the refresh cycle. A model is a snapshot of how your channels behaved, and it goes stale. IAB's 2025 guidance, summarized by Genius Monkey, describes quick weekly refreshes with full retrains monthly or quarterly, and every retrain re-runs the data work above.

What the price list actually looks like

Across the 14 head-term pages I read for this piece, none quoted a price. Fourteen pages explaining a technique, zero numbers attached to it.

The numbers exist. They are published on cost-adjacent pages and vendor pricing pages, by the companies selling the work. Where a figure comes from a vendor's own page it is that vendor's list price. Where it comes from a third party, it is marked as reported.

The gap between $490 a month and $500,000 a year is not a pricing quirk. It is the difference between a self-serve tool that fits a standard model to your spend file, and a program where people build, defend and refresh a custom model for you.

Notice which vendors chose to publish. Cassandra's tier ladder moves with your own media spend from the previous year, which is unusual honesty in this category. Recast, Northbeam, Keen and Mutinex all run custom or enterprise quotes instead, which is ELIYA's observation in the same comparison. That comparison puts the reason plainly, citing Gartner: opaque pricing dominates MMM, and most vendors do not publish prices. That is not a small inconvenience for a buyer. It means two quotes cannot be compared until you force both vendors to scope the same model, the same data work and the same refresh cadence onto paper.

Two more things sit outside every figure in that table. ELIYA notes that most MMM and incrementality providers require annual contracts, and that the price quoted on a sales call rarely includes onboarding fees. The same vendor guide puts the cost of a single incrementality test beyond the platform fee at 5 to 15 percent of the tested channel's spend in forgone sales during the window. That figure belongs to the test rather than the model, and it belongs in the same budget conversation.

The Threshold Is Published. It Is Also Misleading.

The threshold is not a secret. Vendors print it, and Google prints one with the arithmetic attached.

Google's Meridian documentation states the rule plainly: two years of weekly data for geo-level models, three years for national-level models, and three years if you only have monthly data.

So why does a model built on two years of clean weekly data still fail? Because two years is a duration, and duration is not the constraint that binds.

How much history a model actually needs, and why variance beats observation count

Meridian's own documentation answers this, and among everything I read for this piece it is the only source that converts the question from years into a ratio.

Twelve media channels, six controls and eight knots come to 26 parameters in Meridian's own worked example. Two years of weekly data gives you 104 data points. That is four data points per parameter, and the documentation's own verdict is that this sample size is too low to estimate the model reliably. It adds that insufficient variation in media spend makes it worse.

Stretch to three years and a smaller model in the same document, and 156 data points across 10 parameters gives roughly 15 per parameter. At that ratio you get something the documentation calls directional information. Direction, not certainty.

That ratio is the test to run before you buy. Count your channels and controls, count your weeks, divide. Meridian's two examples sit at four and at roughly fifteen, and the distance between them is the difference between a coefficient that survives a serious question and one that does not.

Variance matters as much as count. Recast, a measurement vendor, published the cleanest demonstration of it. In their simulation, a channel whose spend ranged from $200 to $600 was recovered within a few percentage points of error on at least six months of data. A channel with similar variance but under $100 spent per day never showed up as significant. Their own conclusion is the line worth keeping: it is the variance in a channel that matters, not just the number of observations.

A channel can also have the variance and still fail, if its effect is too small to detect. That counter-example is part of the same finding, which is what makes it credible rather than convenient.

The published minimums do not agree with each other, and the disagreement is the story:

Six sources, minimums running from six months to five years. Anyone who tells you the answer is two years and stops there has handed you a number without the condition that makes it true.

The way to read a disagreement like that is to ask what each vendor's own product is shaped like. Measured sells a product that starts at six to twelve months, and its guidance describes traditional MMM at two to three years before describing its own floor. Meridian is Google's own open source framework, so its national rule reflects the kind of model it was built to fit. Neither is lying. They are pricing different shapes of the same problem. Find the source whose model shape matches yours, then run the ratio on your own numbers rather than trusting their headline.

What happens when one channel is 4 percent of spend

This part comes from Google's own documentation rather than from a vendor.

Meridian's guidance says that for media channels with low media spend, it is advisable to combine them with other channels to avoid issues with ROI estimation. Its list of remedies for thin data includes estimating fewer channels, either by dropping a channel with low spend or by combining channels.

The precise wording matters. The documentation does not say the model recommends cutting your small channel. It says a channel too small to estimate gets pooled into another channel or dropped from the model, and either way it does not get its own coefficient.

Which means the budget decision for that channel gets made without the model. The model is not wrong about the channel. It was never asked.

A 4 percent channel is the most exposed line in a media plan for exactly this reason. It is large enough to matter to the business and too small to earn its own number. If it also carries the worst dashboard, it is the first thing the next budget cycle cuts, and the model you just bought will not defend it.

The practical move is to decide the pooling in advance rather than let it happen to you. List every channel you plan to ask the model to separate, mark the ones you expect to fall under the threshold, and decide now how you will defend those budgets if they come back pooled. A channel defended by a measured spend shift stands on its own whether or not the model could see it.

What the Model Settles, and What Only a Test Can

A fitted marketing mix model gives you elasticities and contribution shares, and those are genuinely useful for allocating budget between channels you can already separate. It tells you how sales moved with each channel across the period you fed it, with a range attached.

IAB's State of Data 2026, surveying more than 400 senior decision-makers, found that no respondents believe all paid channels are represented well in current marketing mix models. Nobody in the sample. That is the trade body's finding, not a vendor's, and it should set your expectations before you pay for the first model.

What the model cannot do is confirm its own assumptions. If two channels always rise and fall together, the model will split their credit somehow, and the split will look like a finding. Polestar Analytics, the most thorough page I read on this topic, names the same weakness directly: MMM needs history and struggles with correlated channels. Sigmoid makes the point under its better name, multicollinearity, and adds the consequence that models restricted to limited historical data fail.

One more shape to respect before you trust a coefficient: the response curve it came from. Advertising effect builds and decays, and a model fitted across a period with a promotion in it will attribute the promotion to the channel. Polestar lists curve shape and decay among its own caveats. If your coefficient came from twelve unusual weeks, it describes those twelve weeks.

So the output is a set of claims, not findings. Pick one number and hold it until something outside the model confirms it.

The two questions a coefficient cannot answer

A coefficient tells you how much a channel contributed given what happened. That already assumes the rest of the mix did what it did.

The two questions it cannot answer are the two a budget review actually asks. What would have happened if you had spent nothing there, and which of two channels that always move together deserves the credit. Both need a version of your business that did not happen. Attribution cannot answer them either, because it allocates credit for what did happen. Our piece on revenue attribution covers what that method can and cannot carry.

For this article the distinction collapses to one sentence. The model describes the past you gave it, and the test creates a future you did not have.

The one test that tells the model is wrong

Calibration is small and bounded, and its payoff is large. You take a channel, withhold it somewhere, and compare what the model predicted against what actually happened. If the prediction holds, your coefficients earned some trust. If it does not, you found out before you moved a budget on them.

Strong.io describes the calibration test this way, and it is the correct description. None of the 14 pages I read turns the result into a gate you check before you buy. It stays a validation step described after the modeling, which is the wrong order if the point is to avoid spending.

That gate matters because trust is not evenly distributed. Haus, which sells incrementality testing, publishes a 2026 Marketing Decision Confidence Index reported by EMARKETER: 60 percent of US senior decision-makers trusted independent incrementality testing most, ahead of media mix modeling at 40 percent and in-platform reporting at 37 percent. Your decision-maker is more likely to believe the test than the model. Plan the work in that order.

If you want the designs and the arithmetic, our guide to incrementality testing carries the mechanics. The point here is narrower. A model you have not calibrated is an unverified claim about your own business, and the check is a test you can scope and price before you commit to the program.

Build, Buy, or Skip

Three questions decide this, and none of them is about which vendor.

Can your data support the ratio? Meridian's examples run from four data points per parameter, which it calls too low, to roughly fifteen, where the output becomes directional. If you land near the bottom, you are buying a model that will produce numbers and not much else. Fix the data shape first, or model fewer channels.

Do you have someone who owns the refresh? Not someone who can run it once. Someone for whom the monthly re-fit is a standing task. Odins describes that cycle as continuing for as long as the model runs, and a model without an owner decays into a slide nobody trusts.

Is the decision big enough to justify it? A model earns its keep when you are moving real money between channels you cannot separate. If your budget sits in two channels, a staged spend shift will teach you more per dollar than a full model will.

If all three answers are yes, the open source route is genuinely available. Meridian, Robyn and PyMC cost nothing to license, and the trade is that you supply the data scientist the paid vendors were going to supply for you. If you have that person and the ratio holds, build. If either is missing, buy the managed version and pay for the missing half.

If one answer is no, skip it this year and put the money into evidence you can actually use.

Whichever way you go, buy the check before you buy the program. A calibration run against one withholdable channel is a scoped test rather than a program, and it tells you how a vendor behaves when their number meets your reality. For the buying questions that come before MMM specifically, see marketing attribution software. Once the model exists you still need somewhere to report it, and our round-up of marketing analytics tools covers that layer.

What to Do While You Cannot Afford One

If you are weighing this, you probably will not buy a model this quarter, and that is a reasonable outcome. You can still answer the question the model was supposed to answer, with evidence instead of inference.

Stage the budget shift instead of modeling it. Move a defined amount from one channel to another for a fixed window and watch sales. You are not estimating an elasticity, you are measuring one. The evidence is weaker than a fitted model and it is real, which beats a model you cannot calibrate.

Use geo splits as evidence. Hold a channel in some regions and run it normally in others, then compare. It is the same logic as a withhold without the platform fee, and it is the fastest way to find out whether a channel does anything. The designs are written up in more detail in our incrementality testing guide.

What neither method will tell you is the long-term effect, the brand contribution that shows up two quarters later, or what a channel you have never run would do. Those are the questions a model is actually for, and being clear with your team about which question you are answering saves you from buying a model to solve a test's problem.

The spending is already happening elsewhere. IAB and BWG Global's State of Data 2026, a survey of 430 US senior planning and analytics leaders, found that between 67 and 76 percent currently use incrementality tests, attribution analysis or marketing mix models, with 67 percent naming MMM specifically. The gap between owning one and acting on it is the part worth sitting with: HBR Analytic Services, in a 2026 survey reported through an industry statistics round-up, found that 87 percent say MMM matters to their organization while only 28 percent are very effective at acting on it.

You do not close that gap by buying a bigger model. You close it by having one number you trust before you spend anything on the next one, and one number has to survive the review it lands in. Our guide to marketing KPIs is the place to start on making that number hold up.