Marketing attribution vs MMM vs incrementality: which question are you asking?
Quick answer
Attribution divides up credit for the interactions you can track. Marketing mix modelling helps you assess spending across channels. Incrementality testing asks what changed because you ran the campaign. Choose the method that answers your question; don’t expect three different methods to produce the same number.
Three methods, three different jobs
A marketing team asking why its reports disagree may be comparing methods that were never designed to produce the same answer. A platform can count conversions after ad exposure, a CRM can credit the first recorded source, and a model can estimate the effect of spending across regions. Each describes a different view.
Begin with the decision. Daily campaign troubleshooting needs timely operational data. Annual allocation needs a view across channels and external conditions. Testing whether a campaign creates additional sales needs a counterfactual.
Trying to make one report serve all three purposes tends to produce false certainty. A detailed journey diagram can still be incomplete; a sophisticated aggregate model can still depend on weak assumptions.
Use attribution to understand recorded journeys
Attribution rules distribute credit among observed interactions. First-touch and last-touch rules are easy to explain but deliberately simplify the journey. Data-driven methods can use richer patterns, while still depending on the events and identities available.
Missing consented observations, device changes, offline conversations and untracked sharing can leave gaps. Those gaps do not mean the interactions never happened. They mean the report cannot observe them reliably.
Operationally, attribution is valuable for finding broken source capture, comparing recorded paths and identifying where prospects appear to stall. Combine it with a consistent campaign-tagging system and clear CRM definitions. Avoid relabelling an attribution allocation as a causal experiment.
| Method | Best suited to | Main limitation |
|---|---|---|
| Attribution | Recorded journeys and operational diagnosis | Unobserved interactions and non-causal credit |
| MMM | Broader allocation and scenario analysis | Model assumptions and correlated inputs |
| Experiment | Effect of a defined change | Applies to tested conditions |
| Reconciliation | Consistent reporting definitions | Does not itself establish causal lift |
Use MMM for broader allocation questions
Marketing mix modelling works with aggregate time-series or geographic data to estimate how marketing and other factors relate to outcomes. Inputs may include spend, reach, sales, price, promotions, seasonality and distribution. The quality of the comparison matters more than the number of variables.
Google’s Meridian introduction describes a model using business outcomes, media activity and relevant controls. Its causal-inference documentation makes clear that causal interpretation depends on model structure and assumptions.
A model cannot automatically separate two channels that always rise and fall together. Nor can it reconstruct a major business change that was never recorded. Treat sensitivity checks, uncertainty and external validation as part of the deliverable.
Use experiments to test a defined change
An incrementality experiment compares outcomes with and without a specified activity under a design intended to make the groups comparable. Randomisation is powerful where feasible, but the implementation must account for spillover, incomplete exposure and the time needed for outcomes to appear.
The result applies to the tested audience, budget, creative and period. A positive result at one spending level does not establish unlimited profitable scale. A short test may miss delayed effects.
Experiments can also inform a broader model. Google’s calibration guidance explains how experimental information can inform prior assumptions. That is a useful connection between methods, provided the experiment and model describe compatible conditions.
When the reports disagree
Say your paid social platform reports £120,000 in revenue from £30,000 in spend, but finance can’t see anything like that growth. Before arguing about which report is right, line up the dates and definitions. Check for duplicated conversions, cancellations and revenue that more than one channel has claimed.
Next, use an experiment to estimate the effect of a specific paid-social change. If a credible test estimates £18,000 in incremental contribution for £15,000 of incremental spend, the simple net return on that tested increment is 20%. It is not the same quantity as the platform’s revenue-based ROAS.
Finally, combine the evidence with broader channel and business data for allocation. The marketing warehouse guide explains why consistent record grain and timing matter before those datasets are joined.
Get the numbers to answer a question
Before commissioning more measurement, we’d ask what decision it will change, what data it uses, what’s missing and what assumptions it makes. Then agree what result would make you act. If those answers are vague, a more sophisticated model won’t make the proposal more convincing.
Use attribution for operational visibility, models for allocation under stated assumptions, and experiments for causal learning where feasible. Investigate disagreement rather than averaging incompatible numbers into a comforting compromise.
For a small team, begin with clean definitions and a manageable analytics review. Better measurement is not necessarily another platform; it may be a reconciled dataset and one experiment that answers a decision you have been avoiding.
Related reading
Marketing incrementality testing: find out what the campaign changed · UTM naming conventions: a campaign tracking system people will use · Marketing budget allocation: fund the next useful decision
Photo: Carlos Muza / Unsplash.
