MMM vs. attribution vs. incrementality testing: which to use when

By Mo Touzani · 2026-07-06 · Marketing Measurement

You have three measurement methods competing for the same budget conversation: marketing mix modeling, multi-touch attribution, and incrementality testing. Vendors pitch each one as the answer. They answer different questions. A team that knows which question each method answers builds a measurement stack that supports decisions. A team that treats them as interchangeable buys three dashboards that disagree with each other and picks whichever number supports the plan they already had.

The three questions

Multi-touch attribution answers a descriptive question: which channels appeared in the paths that converted, and in what sequence? MMM answers a modeling question: given the spend history across channels, what allocation does the historical pattern suggest? Incrementality testing answers a causal question: if you removed or added this spend, how much revenue would change?

Only the third question is causal. Attribution observes paths. MMM fits curves to historical spend and revenue. Both describe what happened alongside the spend. Neither can tell you what the spend caused, because neither includes a control condition. Incrementality tests include one by design, which is why they cost more to run and why their answers hold up when budget decisions get challenged.

This ordering does not make attribution and MMM useless. It makes them tools with a defined scope. The failures come from scope creep: an attribution report used to justify cutting a channel, or an MMM curve used to promise a board that doubling spend will double return.

Where multi-touch attribution earns its keep

Attribution data shows you touchpoint sequences, frequency patterns, and the funnel stages where buyers stall. Use it to answer operational questions: is the retargeting frequency cap too loose, are buyers hitting the pricing page before or after the case studies, which entry pages feed the demo funnel. These are path questions, and attribution is a path instrument.

Attribution breaks when you ask it to rank channels by causal contribution. Bottom-of-funnel channels collect credit for buyers who had already decided. Brand channels diffuse their effect across weeks and get undercounted. The ranking that comes out reflects proximity to the conversion event, and proximity is a property of funnel position rather than of causal power.

Signal loss since iOS 14.5 has degraded path data further, and the platforms now fill gaps with modeled conversions. Treat attribution numbers as a sample of observable paths, weighted toward the buyers your tracking can still see.

Where MMM earns its keep

MMM works from aggregate data: weekly spend by channel, revenue, and context variables like seasonality and promotions. It needs no user-level tracking, which makes it durable against privacy changes, and it can cover channels attribution never sees, like linear TV, audio, and out-of-home. For a company spending across eight or more channels, MMM gives the only single view of the whole portfolio.

The limits are structural. MMM learns from the variation in your spend history. If you never turned a channel off, never moved its budget by more than 15 percent, the model has no evidence about what happens outside that band. It will still produce a number for you. The number is an extrapolation. Newer architectures produce tighter confidence intervals around these extrapolations, and the tighter intervals measure fit to history rather than knowledge of counterfactuals.

An MMM trained on observational spend history can tell you what your past allocation correlated with. It cannot tell you what a different allocation would have produced. Only an experiment can.

Where incrementality testing earns its keep

An incrementality test withholds a channel from a randomized control group, usually defined geographically, and measures the revenue difference against regions where the channel keeps running. The output is causal: incremental revenue per dollar, with a confidence interval, for the channel at the tested spend level. When a test result and an attribution report disagree, the test is right.

The costs are real. A test occupies a channel for four to eight weeks, requires enough conversion volume in the control regions to detect the effect size that matters, and answers one question per test. You cannot test every channel every quarter. A company running a serious program gets through four to six tests a year, which means the testing calendar is a prioritization exercise, and the channels most likely to be over-credited by attribution, retargeting, brand search, and affiliates, go first.

Choosing by decision, budget, and stage

Below roughly $50,000 per month in paid media, skip MMM. The model needs spend variation across channels to learn from, and at that scale you have neither the channel count nor the data history to feed it. Run platform attribution for operational path questions, keep a clean internal baseline by traffic source, and run your first geo-holdout test on the channel whose value you doubt most. One test at that scale beats any model.

Between $50,000 and $500,000 per month, add a lightweight MMM, open-source implementations are adequate here, and run a standing test calendar of two to four incrementality tests per year. Use the test results to calibrate the MMM: when the model says a channel returns 0.7 and the geo test says 0.3, adjust the model priors rather than averaging the two numbers.

Above $500,000 per month, all three methods run continuously and the design question becomes integration. Tests feed calibration priors into the MMM on a set schedule. Attribution stays scoped to path diagnostics and creative-level operations. Budget authority sits with the calibrated model, and no reallocation above a defined threshold ships without either a supporting test result or a scheduled test to validate it.

The framework for building the stack

1. Assign each method a question it is allowed to answer, in writing. Attribution answers path and sequence questions. MMM answers portfolio allocation questions at the tested spend range. Incrementality tests answer causal questions about specific channels. Write the scope down and enforce it in budget reviews. Most measurement dysfunction traces to a method answering a question outside its scope, and a written scope makes the violation visible.
2. Sequence the build: baseline, then test, then model. Start with internal conversion baselines by traffic source, since every later method gets evaluated against them. Add geo-holdout testing next, because a single clean test produces causal evidence no model can. Add MMM last, once you have test results to calibrate it against. Teams that buy the model first end up with confident output and no way to check it.
3. Route every disagreement to the most causal method available. When attribution and MMM disagree, the MMM view wins for portfolio questions and attribution wins for path questions. When either disagrees with a well-designed incrementality test, the test wins. Adopt this precedence rule before the disagreement arises. Deciding it in the moment, with a budget on the line, produces whichever answer the loudest stakeholder prefers.

What the integrated stack buys you

A measurement stack with clear scopes and a precedence rule changes the texture of budget meetings. Channel ROI claims arrive with their evidence class attached: observed path, modeled estimate, or tested effect. Reallocation proposals name the test that supports them or the test that will check them. Disagreements between methods become calibration inputs instead of political ammunition.

None of this requires platform-scale infrastructure. It requires deciding which question you are asking before you pick the tool that answers it, and holding that line when a dashboard offers a convenient number for a decision it cannot support.