Google gave away the model. That was never the expensive part.
Google open sourced Meridian. Meta open sourced Robyn. PyMC Marketing has been free the whole time. If the model was the product, this category should have collapsed. It did not, and the reason is worth understanding whether you are buying media mix modeling or building it.
The giveaway
A Bayesian media mix model is a well understood object. You need a likelihood over your outcome, adstock to carry effects past the flight, a saturation curve so returns diminish, controls for price and seasonality and whatever else moves your category, and a sampler to fit it. All of that is public. Google and Meta both published production quality implementations, and the literature behind them goes back decades.
Fitting one is fast. On two or three years of weekly data across a dozen channels, a modern sampler finishes in minutes on a laptop. With a competent analyst and clean data, you can go from CSV to posterior in an afternoon.
So why does an engagement cost six figures and take a quarter?
Where the time actually goes
Here is the honest breakdown of a twelve week engagement, from someone who has run them.
- Scoping and alignment, one to two weeks. Deciding what outcome you are even modeling. Revenue, contribution margin, and new customer acquisition give different answers and imply different budgets.
- Data collection and reconciliation, four to six weeks. The largest block by a wide margin. Pulling spend from every platform and agency, mapping thousands of campaign names to a channel taxonomy somebody has to invent, reconciling three systems that disagree about what a conversion is.
- Modeling, two to three weeks. The part everyone thinks they are buying. Mostly iteration on specification rather than compute.
- Validation, one to two weeks. Holdout error, and if you are fortunate, comparison against an experiment that actually ran.
- Socialization, two to four weeks. Explaining to the paid search team why the model credits them with less than their platform dashboard does.
The free model addresses the third item. It is roughly a fifth of the work, and it is the only part that was ever going to be commoditized.
The model was never the moat. Running it was.
What stays hard
The data pipeline. Every advertiser has a different mess. Agency reporting arrives as spreadsheets with inconsistent naming. Platform APIs change. Offline sales land weeks late and get restated. None of this is intellectually interesting, which is exactly why it stays expensive.
Specification. Meridian will happily fit whatever you hand it. It will not tell you that your branded search variable is absorbing the effect of every upper funnel channel, or that your promotional calendar is collinear with your television flights. Those judgments determine the answer, and getting them wrong produces a model that fits beautifully and misallocates budget.
Identification. This is the one nobody advertises. A meaningful share of media mix models are fit on data that cannot support the causal claim being made of it. Spend never varied enough. Two channels always moved together. No holdout was ever run. The model returns a number regardless, with a credible interval that looks reassuringly tight, and a consultant who has already invoiced has limited incentive to say so out loud.
Calibration. A model checked only against its own fit is checked against nothing. Holding estimates to measured lift from a geo experiment is what separates a defensible number from a plausible one, and it requires running experiments, which requires cooperation from people whose job is spending the budget rather than measuring it.
Cadence. Because every refresh repeats most of the above by hand, refreshes get rationed. Quarterly is not a methodological choice. It is a pricing artifact. The data would support weekly.
What this means if you are buying
Ask what you are paying for. If a vendor is charging enterprise rates for a wrapper around an open source model, that is worth knowing. If they are charging for pipeline engineering, experimental calibration, and the judgment to tell you when your data cannot answer your question, that is a different proposition and probably a fair one.
Two questions worth asking in any evaluation.
- What would make you tell me this model is not trustworthy? A good answer is specific: insufficient spend variation, collinear channels, no experimental validation. A vague answer means nobody checked.
- How often can this be refreshed, and what does each refresh cost? If the answer is quarterly and another engagement, you are buying a report. If it is weekly and compute, you are buying a system.
What this means if you are building
Start from Meridian or PyMC Marketing. Writing your own sampler is a way to spend six months arriving where you could have started. Put the effort into ingestion, taxonomy mapping, specification search with honest out of sample selection, and automated diagnostics that refuse to publish a result when identification is weak.
That last one matters more than it sounds. Automation is only safe where validation is objective. Holdout error, calibration against measured lift, and simulation based calibration are numbers rather than opinions, which is what makes it possible to let a model run unattended and still trust what comes out of it.
We build media mix models that run continuously rather than quarterly, for advertisers between five and fifty million in annual media. If that is you, get in touch.