Media Mix Modeling vs Multi-Touch Attribution: Which Question Can Each Answer?

Media mix modeling estimates how aggregate media and non-media variables relate to a business outcome across time and often geography. Multi-touch attribution assigns credit across observed user-level touchpoint paths. MMM is built for portfolio response and budget questions; MTA is built for finer path-credit questions. Neither method earns a causal claim merely by producing channel percentages.

Google Meridian’s data guidance describes a typical MMM dataset as aggregate media execution and spend, a summable KPI, controls, time, and ideally geography. Google’s MTA evaluation research describes multi-touch attribution as working from observational, user-matched paths of marketing events and outcomes.

That difference in unit creates different questions:

• MMM: “How did the outcome respond to the media portfolio, after modeling carryover, saturation, controls, and uncertainty?”
• MTA: “How should credit be distributed across the recorded touches on matched paths before a recorded conversion?”

There is no single universal formula for either method. MMM is a family of statistical specifications; MTA includes rules and fitted models. A result is meaningful only with its data contract, assumptions, version, and estimand—the precise effect or credit quantity the method is trying to estimate.

MMM sees the portfolio through aggregate variation#

An MMM relates changes in a KPI to changes in media and relevant non-media variables. Meridian’s causal-methodology overview frames MMM as a causal-inference problem using aggregate observational data. Its model specification includes choices for media carryover, nonlinear response, geography, trends, seasonality, and non-media treatments.

Those choices let MMM address questions such as:

• What is the estimated incremental contribution of each modeled channel over the observed period?
• How does estimated response change as spend changes?
• What allocation is consistent with the fitted response curves and constraints?
• How uncertain are those estimates?
• How do results change when priors, controls, time windows, or model structures change?

MMM can include channels that lack user-level identity because the unit is aggregated. It can also incorporate offline media and non-media variables when suitable time and geographic data exist.

Its limits follow the same design. Channels that move together are hard to separate. Missing confounders can bias effects. Weak variation limits identification. Aggregation can hide campaign, audience, and creative differences. A response curve learned in one market period is not automatically stable in another.

MTA sees the path that the identity system recorded#

MTA begins with a different object: a sequence of observed touches associated with an observed conversion. A simple rule might credit the first or last touch. A multi-touch rule can split credit evenly or by position. A fitted model can use differences among converting and nonconverting paths to allocate credit.

This granularity helps with questions such as:

• Which recorded channels and touch types appear on matched conversion paths?
• How does assigned credit vary by campaign, creative, or audience under this model?
• Which path features distinguish the observed converting and nonconverting samples?
• How do credit allocations change when the lookback window or identity rule changes?

The word recorded is essential. Identity may break across devices, browsers, people in a buying group, logged-out sessions, offline interactions, and restricted platforms. In B2B, one person’s content view may precede another person’s procurement action. A user-level path is not necessarily an account-level decision path.

Google Research’s path-evaluation paper treats MTA as an observational user-path problem and evaluates attribution against matched path evidence. That supports careful model evaluation; it does not make every assigned share causal.

Compare the methods by decision, not by sophistication#

“Better” means better fit to the question and available evidence, not more advanced mathematics.

Prediction fit does not validate causal effect#

Meridian’s model-fit guidance states the problem directly: the goal is causal inference, but causal quality is difficult to validate without well-designed experiments. A model can predict held-out KPI values while assigning the wrong causal explanation if confounding and model structure are wrong.

That creates a minimum review contract for MMM:

• a causal graph or written mechanism for treatments, controls, and confounders;
• the aggregation, time range, geographic unit, and KPI definition;
• media exposure and spend definitions;
• priors and experiment calibration, if used;
• sensitivity to controls, windows, and model forms;
• posterior uncertainty and decision implications;
• a record of data and model versions.

MTA needs a parallel contract:

• identity unit and match coverage;
• included channels and unavailable surfaces;
• conversion definition and deduplication;
• lookback window and path construction;
• treatment of direct, organic, offline, and cross-device touches;
• credit rule or fitted model;
• evaluation sample, comparator, and version.

Without those fields, two dashboards can disagree for reasons that are invisible to the decision-maker.

Do not force the outputs to match#

MMM may estimate a channel’s portfolio contribution even when few user-level paths are available. MTA may allocate detailed credit inside a channel that MMM cannot distinguish confidently from another channel. Different time windows, KPIs, geographies, and identity units can legitimately produce different shares.

Reconciliation should start with definitions:

1. Are both methods measuring the same outcome?
2. Do their time windows and attribution lags align?
3. Are channels grouped the same way?
4. Does MTA omit exposure that MMM includes?
5. Does MMM aggregate campaign differences that MTA exposes?
6. Is either output being described as causal without supporting design?

Only after that should the team investigate model error.

Use a measurement portfolio#

The strongest operating model gives each method a bounded job. MMM informs portfolio allocation and response scenarios. Path analysis and MTA inspect recorded journey detail. Experiments calibrate specific effects when practical. Finance and sales records anchor the outcome. Qualitative research explains mechanisms and missing context.

No universal history length, spend floor, channel count, fit statistic, or accuracy threshold can replace that decision contract. Data sufficiency is a property of the variation, model, and question.