Unified Marketing Measurement, Minus the Vendor Definition

Unified marketing measurement is the label for putting marketing mix modelling, multi-touch attribution and experiments into one view, on the theory that each covers the others' blind spots. The theory is sound and the usual implementation is not, because three methods that disagree do not average into truth. They average into a confident number with no design behind it, which is worse than any of the three on its own.

The three, and what each actually observes

Marketing mix modelling works on aggregated time series: spend by channel by period against revenue by period. It sees the whole business including offline and brand effects, it needs no identifiers, and it cannot resolve an individual journey. Its weakness is that it is fitted on observational data, so a channel whose spend rose when demand rose can look causal when it was only correlated. Marketing mix modelling versus causal attribution covers the comparison.

Multi-touch attribution works on individual journeys and divides credit among observed touchpoints. It resolves detail the model cannot, and it only sees what the tracking resolved, so it is blind to the unresolvable share entirely. How multi-touch attribution works covers the mechanics.

Experiments withhold a channel from a randomly chosen group and compare. This is the only one of the three that observes a counterfactual directly.

Why averaging them fails

They are not three estimates of one quantity. Marketing mix modelling estimates a channel's aggregate contribution over a period. Multi-touch attribution estimates a share of credit among resolved journeys. An experiment estimates the effect of withholding on a specific group over a specific window. Averaging those is averaging a temperature, a distance and a weight.

Worse, two of the three share a blind spot. Both the model and the journey data are fitted on what was observed, so both inherit the same selection problem: spend went where demand already was. When they agree, that agreement is often the shared bias speaking, not corroboration. The book's account of the Facebook experiments is the sharp version of this, where observational methods overstated effects against randomised ground truth by roughly three times, and six of fourteen comparisons could not be distinguished from zero.

What a defensible unified view looks like

Show the three separately, never blended, and give each one four properties: where the number came from, the coverage of the data underneath it, the design (experimental, quasi-experimental or observational), and the interval.

Run at least one randomised holdout per quarter on the channel carrying the most budget, and use it to calibrate the observational components. When the model and the journey data both disagree with the experiment, the experiment wins and the other two need re-fitting.

Where a causal read fits

Between anchors. A causal read estimates each channel's incremental contribution from the variation already in the data. It is not a substitute for the experiment and it does not claim to be.

What to do this week

• If you have to defend the number: stop reporting a blended figure. Report three, each with its design named. The disagreement is information your CFO should see.
• If you own the budget: check when your last randomised holdout ran. If the answer is never, your unified view has no anchor and is three observational methods agreeing about the same gap.

Why unified measurement breaks when the three systems disagree covers the reconciliation problem.

As of 9 September 2026. The Facebook experiment comparisons and the four properties are from The Price of Being Found (Edition 2.10), Chapters 12 and 15, with the book's caveats.