Media mix modeling
From free-plus-a-data-scientist to six-figure engagements, and why the invoice is rarely the real price.
01 / The answer in one paragraph
Media mix modeling costs anywhere from nothing to several hundred thousand dollars a year, and the spread comes from one variable: who does the work. Open-source libraries are free and need a data scientist. Measurement platforms bundle the data scientist into a subscription. Consultancies sell the work as a project. SignalRig packages the work into a $249 Mac app for teams that cannot justify any of the above. Details and caveats below.
02 / The four price tiers
| Tier | Typical cost | What is included | What is not |
|---|---|---|---|
| Open source | $0 for the software | Robyn, Meridian, or PyMC-Marketing. Full model code, documentation, community. | The analyst. Budget a part-time data scientist or a contractor at market rates, and their time every refresh cycle. |
| Consultancy project | Five to six figures per engagement | Data prep, modeling, a presentation, expert narration. | Refreshes. Every new quarter is a new fee, and the model usually stays in their environment. |
| Measurement platform | Subscription, commonly low thousands per month and up, annual contract | Hosted modeling, dashboards, support, sometimes incrementality tests. | The model itself, in most cases. You see outputs, not the exam. Your data lives in their cloud. |
| SignalRig for Mac | $249 once | The fitting engine, the strict data validator, the graded exam (planted-truth recovery and holdout accuracy), plain-language verdicts, media-plan export. Model core is open source. | Managed service, multi-brand automation, an API. It runs on one Mac and works from your own ad platform exports. |
Platform and consultancy figures are ranges drawn from publicly discussed pricing as of 2026; most vendors quote per contract and do not publish list prices.
03 / The costs nobody itemizes
Every option above assumes weekly spend by channel and a weekly outcome series, aligned, complete, and long enough. Getting there is where most first attempts die. SignalRig front-loads this: its validator reads your CSVs and returns specific fixes (a missing week, a channel that is zero for the whole window, an outcome column that is not numeric) before any modeling happens. That is not free either, but it is an afternoon instead of a statement of work.
A model that hands you a number with no interval and no exam is cheap to buy and expensive to act on. The question to ask any vendor is simple: on data where the true answer was planted in advance, how many parameters did the model recover inside its own stated ranges, and how did it do on weeks it never saw? If nobody can answer, the price on the invoice is not the real price.
A mix model is a photograph, not a feed. Plan on refitting quarterly at minimum. Subscription pricing accounts for that; project pricing does not; a one-time app does it for free but on your own clock.
04 / When each tier is the right call
05 / Next