Skip to content
Measurement8 min read

Marketing mix modeling can't tell you when it's guessing

Every marketing mix model converges on an answer, including the ones built on data with nowhere near enough variation to support one. The model doesn't know that, and most vendors selling it have no incentive to tell you either.

The model doesn't know it's guessing

Marketing mix modeling regresses total outcomes — revenue, conversions — against spend and other variables across time, and estimates a coefficient for each channel's contribution. Sold correctly, this is genuinely the strongest available answer to the channel-attribution question at scale: it's causal in ambition, not just correlational like last-click or data-driven attribution.

Sold incorrectly — which is most of the time — it's presented as reliable regardless of the data feeding it. A regression will produce a coefficient for every channel in the model whether or not that channel had enough independent spend variation to be statistically separable from the others. Nothing in the output flags that it's guessing. The number looks exactly as confident on thin data as it does on strong data.

That asymmetry matters commercially. Nobody who sells MMM services benefits from telling a prospective client their spend doesn't support the model — the sales conversation ends. So the disqualifying math tends to live in data-science blogs written for analysts, not on the pages a marketer or founder actually reads before commissioning one.

The threshold that actually matters

The number usually quoted — 'you need two to three years of history' — is real but secondary. History length matters because it gives the model more cycles to observe, but a business can have five years of history and still fail the check that actually decides whether the output is trustworthy: weekly spend variation per channel.

  • Why variation, not volume, is the requirement

    A regression separates one channel's effect from another's by watching them move independently over time. If two channels have moved in lockstep — scaled up together for the same product launches, cut together in the same slow quarters — the model has no way to tell which one actually drove the outcome. It sees correlated inputs and has to guess how to split credit.

  • The practical floor

    Below roughly $10-15k a week in a given channel, most businesses simply don't generate enough independent, meaningful variation in that channel's spend for the model to isolate its effect with any confidence. The model still produces a coefficient. It just tends to land close to whatever a naive spend-proportional split would have given you anyway.

  • What that means in practice

    If MMM's answer for your smaller channels is suspiciously close to 'each channel gets credit roughly proportional to what you spent on it,' that's not a coincidence — it's what the model defaults to when it can't find real signal to work with.

The two assumptions doing the real work

Two structural choices sit underneath every MMM output, and both are commonly left at whatever a vendor's software defaults to rather than fit to the actual business.

  • Adstock decay

    How long an impression's effect persists after the impression happens — a purchase three weeks after seeing an ad might still be partly attributable to it. Get the decay curve wrong and the model either credits spend for outcomes it didn't cause, or misses delayed effects entirely.

  • Saturation curves

    The point at which additional spend in a channel stops producing proportional returns. A model without a realistic saturation curve will recommend pouring more budget into whatever channel currently looks efficient, long past the point where it actually is.

  • Why defaults are the quiet failure mode

    Vendor software ships with default decay and saturation parameters because every implementation needs some starting point. The failure isn't using defaults — it's never revisiting them against the business's own conversion-cycle length and channel-specific saturation behavior, which is exactly the validation step that gets skipped under deadline pressure.

Ask any MMM vendor directly: what's our weekly spend threshold per channel for a trustworthy coefficient, and which of our channels currently fail it? A vendor who can answer specifically is doing the work. One who says 'the model handles that' usually isn't.

Where this leaves smaller advertisers

None of this means MMM is a bad method — at genuine scale, with real spend variation across channels and enough history to observe multiple cycles, it's the strongest tool available for the channel-allocation question. The failure mode is buying it below the threshold where it can actually do that job.

Below that threshold, a geo holdout test on your one or two largest channels gets you a real, defensible causal answer for a fraction of the cost and complexity — see our piece on running an incrementality test for the mechanics. It answers a narrower question than MMM promises, but it answers it honestly, which is the trade that actually matters at this spend level.

The businesses best served by MMM tend to already sense it — enough channels, enough spend, enough history that the sales pitch and the actual math agree. Below that, the discipline isn't refusing measurement. It's choosing the measurement method your data can actually support, rather than the one that sounds most sophisticated on a slide.

We do this for a living

If you'd rather not build this yourself, these are the services where it lives.

Rather not DIY

We'll run this on your account.

A free 30-minute teardown of your funnel, ads, and analytics. You keep the findings whether or not we work together.