What is marketing mix modeling?

Every marketing channel reports its own success, and the sum of those reports is always more revenue than you earned. Platform dashboards count conversions they touched, not conversions they caused. Marketing mix modeling exists to answer the harder question: if a channel's budget had been zero, how much revenue would have disappeared? That quantity is incrementality, the sales that would not have happened without the spend.

The method is a regression, a statistical fit between historical inputs and revenue, over aggregate data: weekly spend per channel, weekly revenue, plus the non-marketing drivers a finance team would insist on (price changes, promotions, seasonality, distribution). It deliberately ignores individual users. No pixels, no cookies, no device IDs. This made MMM look old-fashioned in 2015 and made it indispensable after Apple's App Tracking Transparency framework made user-level tracking opt-in.¹

A model that fits well decomposes revenue into a base, revenue attributable to brand equity, distribution, and demand that exists without marketing, and an incremental slice per channel. That is the waterfall at the top of this page. Everything else an MMM produces (response curves, budget optimizers, ROI tables) is derived from that decomposition.

Contribution decomposition. The share of revenue each channel caused last period. This is the slide that ends the "does TV do anything" argument, in either direction.

Response curves. For each channel, a curve mapping spend to expected revenue. The curves are concave: diminishing returns mean the second million buys less than the first. The bend of each curve, its saturation, is where reallocation decisions come from. Our curve lab lets you drag the parameters yourself.

Optimal allocation. Given the curves, an optimizer proposes the budget split that maximizes modeled revenue. Treat it as a proposal, not an instruction; a model that has never seen you spend $5M on a channel has no earned opinion about what $5M would do. The full output list, with the caveats each item deserves, is on the outputs page.

The requirements are unglamorous and non-negotiable: roughly 104 weeks of weekly history as a practical floor, spend across four or more channels, genuine variation in that spend, and one finance-grade revenue series as the dependent variable, the number the model explains. Geo-level splits help more than most buyers expect; modeling fifty regions at once is like running the same experiment fifty times, and it is why Google's Meridian framework is built geo-first.³

If you cannot meet that list, the honest move is not a worse model; it is a different method. Geo experiments answer big single questions at lower spend. The readiness checklist turns this into six yes/no questions, and when not to use MMM says the quiet part in full sentences.

Three serious open-source frameworks exist, all license-free: Google's Meridian (Bayesian, geo-first), Meta's Robyn (ridge regression with evolutionary search), and the community library PyMC-Marketing.³ Free means the license; the operating cost is the person. The BLS puts the median US data-scientist salary at $120,230², and employer cost data implies roughly 1.46x⁵ that once benefits load in.

The managed market splits into vendors that publish pricing and vendors that quote it. Published, as of August 2026: Sellforte from $2,990/mo⁴, Cassandra from EUR 2,700/mo⁶, Paramark from $100,000/yr⁷. The other nine managed vendors in our registry publish nothing, and unlike most pages ranking for this query, we do not invent numbers to fill the gap. Every "best MMM software" list in this SERP is written by a vendor that ranks itself first; this site sells no MMM and ranks nobody first.

The full decision path is the decide shelf: readiness, then the build-vs-buy calculator, then the cost page, then the 12 questions for the AE call.