Marketing mix modelling vs attribution: UK retailer's comparison

What to take away

Attribution links conversions to touchpoints and is useful for digital journey diagnostics, not for total budget allocation.
Marketing mix modelling uses aggregate sales and media data to estimate media effectiveness across channels, including retail media.
Incrementality testing supplies causal evidence; without it, both methods describe correlation and require caution.
The ICO expects a lawful basis for personal data processing, so identifier-heavy attribution needs a documented check.
A UK retailer should run one decision table before buying either method, with the budget question stated first.

Where the two methods differ

Attribution and marketing mix modelling answer different questions. MMM is a form of econometric modelling. Attribution reads user-level or session-level paths and assigns credit to touchpoints. MMM works with aggregated time series, so it can include media with no click path, such as TV or in-store retail media. The marketing mix modelling overview describes it as a technique for estimating the impact of media and marketing inputs on sales.

Attribution vs MMM

Attribution

Unit of analysis: User or session
Main input: Event and touchpoint data
Strength: Journey detail
Main limit: Identifier loss and correlation
Typical use: Campaign diagnostics

Marketing mix modelling

Unit of analysis: Aggregate period
Main input: Sales and media spend
Strength: Channel coverage
Main limit: Model assumptions
Typical use: Budget allocation

Decision question	Attribution	Marketing mix modelling
Unit of analysis	User or session	Aggregate period
Main input	Event and touchpoint data	Sales and media spend
Strength	Journey detail	Channel coverage
Main limit	Identifier loss and correlation	Model assumptions
Typical use	Campaign diagnostics	Budget allocation

Neither method proves incrementality on its own. For a practical comparison of test design, see incrementality testing vs attribution, which keeps the causal question separate from credit allocation.

What ICO expects from statistical models

Attribution often processes personal data, including device identifiers and online identifiers. The lawful basis guidance requires a documented basis before processing starts. Data minimisation and purpose limitation also apply, even when the model output is aggregate.

MMM usually uses aggregated sales and spend data, so it can sit outside direct marketing rules. That advantage disappears if the inputs are built from individual-level identifiers or if the model feeds personalised targeting. Keep the data map and the legal basis aligned. For a row-by-row method, mapping personal data and device tracking in an attribution setup shows how to record those decisions.

How UK retail media changes the choice

Retail media measurement mixes on-site sponsored placements, off-site audiences and store sales. Attribution can see the click or impression path when identifiers survive. MMM can connect the spend to total sales, but it may not separate a retail media halo from broader brand demand.

The ONS retail industry data is a sensible starting point for retail context, but it will not settle a channel-level budget. Treat it as background, then set your own measurement window and control group.

A decision framework for UK retailers

Use this sequence before committing budget:

Five-step decision framework

Write budget decision in one sentence
List dataconsent, identifiers, sales coverage
Choose attribution or MMM, not both
Add incrementality test where material
Record assumptions, review date, stop owner

Write the budget decision in one sentence, with the channel and period named.
List the data you hold, including consent state, identifier type and sales coverage.
Choose attribution for journey diagnostics or MMM for allocation, not both for the same question.
Add an incrementality test where the budget decision is material and a control is feasible.
Record the assumptions, the review date and the person who can stop the spend.

The framework is deliberately plain. If the decision needs weekly channel credit, attribution may be enough. If the decision needs annual budget shifts, MMM plus a test is usually the stronger route.

Where the choice goes wrong

The common error is to compare attribution and MMM on the same output field, such as return on ad spend. Attribution may report credited revenue, while MMM estimates contribution after other factors. The two numbers are not substitutes.

A second error is to treat retail media as a single channel. On-site sponsored products, off-site display and in-store media need separate rows, separate windows and separate control assumptions. Otherwise the comparison hides the effect you are trying to measure.

What evidence to demand from a supplier

Ask for the model specification, data inputs, transformation steps and validation period. Attribution vendors should show how they treat missing identifiers, view-through windows and consented versus modelled users. MMM suppliers should show how they handled seasonality, price changes and promotions.

Do not accept a single contribution percentage as proof of incrementality. Ask for confidence intervals, holdout results or a planned test. Where a supplier refuses, the budget decision is not ready.

Common questions

Is marketing mix modelling better than attribution?

Neither is better in general. Attribution explains digital journeys; MMM estimates media effectiveness across channels. Use the method that matches the decision, then test incrementality where budget is material.

Can a UK retailer use both methods?

Yes. Run MMM for quarterly allocation and attribution for campaign diagnostics. Keep the definitions and time windows separate so the two reports do not contradict each other.

What does the ICO require for attribution models?

The ICO requires a lawful basis for personal data processing, plus data minimisation and transparency. Statistical modelling does not remove those duties if identifiers are processed.

Do we need an incrementality test?

Where the spend decision is large or repeatable, yes. A test gives causal evidence that attribution and MMM cannot provide alone.