Marketing attribution: MTA, MMM and incrementality

What does marketing attribution tell you?

Marketing attribution tells you how a reporting system distributes credit for a measured outcome. Its answer depends on the interactions it can observe, the rules it applies and the window it considers.

Suppose a hypothetical buyer watches a product demonstration, searches for alternatives, reads an email and finally returns directly to purchase. Different models can allocate credit differently while describing the same purchase. None of those accounting choices changes the amount the business received.

The problem becomes serious when teams add the conversion claims of several platforms together. Each platform can apply its own windows, identity methods and eligible interactions. Their totals are therefore not automatically separate groups of customers.

Start with a business outcome ledger: fulfilled orders, recognised revenue or accepted opportunities. Attribution is an analytical view of those outcomes. It should not quietly create a second version of the business.

Why are last-click, MTA, MMM and lift different?

They differ because they observe different units and make different assumptions. Matching the method to the question is more useful than declaring one model the winner.

Google Analytics' attribution guide, checked in October 2026, distinguishes rule-based and data-driven credit allocation. Its documentation also explains how direct visits are handled in the available attribution models. “Last click” therefore needs a precise product definition, not just a familiar label.

Multi-touch attribution, or MTA, distributes credit across observed interactions. Its weakness is incomplete observation: people change devices, decline tracking, buy offline or encounter media that never appears in a path. A detailed model can still be a detailed description of a partial record.

Marketing mix modelling, or MMM, relates aggregate outcomes to marketing and other explanatory variables. Google's Meridian overview, checked in October 2026, presents a Bayesian modelling workflow with diagnostics and outcome analysis. MMM does not require reconstructing every individual's journey, but it does require credible inputs, useful variation and defensible assumptions.

Incrementality experiments compare outcomes under different treatment conditions. Their aim is a causal question: what changed because an intervention happened? Randomisation helps, but broken assignment, spillover or missing outcomes can still undermine a test.

How do you choose the right measurement method?

Choose the method by writing the decision first, then checking whether the available evidence can support it. Avoid buying a sophisticated model before agreeing what action it should inform.

1. Separate diagnosis from allocation

A campaign manager asking why checkout completions fell needs event and journey diagnostics. A finance leader deciding next quarter's channel allocation needs evidence about marginal outcomes and uncertainty. The same chart rarely serves both needs equally well.

Write the proposed action explicitly: pause a campaign, change creative, expand a channel or shift the annual budget. Also name the cost of a wrong decision. More consequential choices deserve stronger evidence.

2. Reconcile the outcome

Agree the definition, time zone, currency treatment, refund window and reporting grain. Distinguish order date from ad interaction date and customer acquisition from repeat purchases. Many attribution disputes are data-definition disputes wearing statistical language.

In a lead business, connect accepted opportunities and revenue back to the enquiry where permitted and feasible. Do not treat a missing connection as evidence that marketing had no effect. Mark it as missing information.

3. Inspect what each method cannot see

List missing offline sales, consent-related gaps, duplicate events and channels without reliable exposure data. Record where spend changed together with pricing, distribution or seasonality. These conditions affect the interpretation of both path models and aggregate models.

For MMM, ask the modeller which variables are plausible confounders, how correlated channels are handled and whether the results are sensitive to reasonable assumptions. A narrow interval is not reassuring if it rests on a misspecified model.

4. Design experiments around decisions

Decide the eligible population, assignment method, outcome, duration and stopping rule before launch. Check whether control units can still encounter the treatment and whether the business can honour the assignment operationally.

Microsoft Research's work on telemetry loss, published at CIKM in 2018 and posted to arXiv in 2019, examines how missing measurement can bias online experiments. Its lesson for marketers is to test the measurement system as carefully as the advertising treatment.

Do not read a non-significant result as proof of no effect. It may mean the test cannot distinguish a commercially useful effect from noise. Report the range of plausible outcomes and what another test would cost.

5. Combine evidence without hiding disagreements

Use attribution for observed journey diagnosis, experiments for selected causal questions and MMM where aggregate planning conditions justify it. Meridian's current documentation describes calibrating models with experiment evidence. Calibration can strengthen a model; it cannot make an irrelevant or poorly executed experiment appropriate.

When methods disagree, examine their populations, time windows and estimands: the exact quantities being estimated. A short promotion test and a long-run model may not be answering the same question.

Where do attribution errors become expensive in India?

They become expensive where purchases cross online and offline systems. Dealer networks, assisted sales, branch enquiries and marketplace purchases can leave an incomplete digital trail.

Imagine an appliance brand whose advertisements create enquiries but whose transactions happen at local dealers. A website-only report can undervalue that activity. Equally, treating every dealer sale near a campaign as advertising-driven would overstate the evidence.

A useful first step is consistent geography, time period and product identifiers across available datasets. Start with lawful, proportionate measurement and an honest coverage statement. More personal data is not automatically a better answer.

Language and regional promotions also matter. If a festival offer runs only in the treatment region, a geographic comparison may estimate the combined package, not media alone. State that clearly rather than labelling the result an advertising effect.

When should a team invest in MMM?

A team should consider MMM when recurring allocation decisions, sufficient historical data and meaningful variation justify the effort. There is no universal spend level at which modelling suddenly becomes valid.

In week one, audit the outcome and spend histories. Next, assess missingness, structural breaks and channel overlap with a qualified analyst. Commission a model only after the data can support the intended question.

Review a model when the business changes materially: a new distribution model, pricing structure or product mix can weaken old relationships. Calendar-based refreshing alone is not a substitute for checking whether the model still describes the business.

Who owns measurement decisions?

Marketing and finance should jointly own the decision, while analysts own the methodological explanation. Platform specialists contribute delivery context, but should not be the sole judges of their platform's causal impact.

Require a short written recommendation with assumptions, uncertainty, alternative explanations and the action supported. Give someone permission to say that the data is insufficient. That is a valid analytical outcome.

My view

I would prefer a modest, well-understood measurement system to an impressive model that nobody can challenge. The important question is whether the recommendation survives reasonable alternative explanations.

The trade-off is that credible measurement can slow a decision and still leave uncertainty. Set a proportionate evidence standard: an exploratory creative test and a major annual budget shift should not require identical proof.

A channel's share of credit is not automatically its share of cause.

Start with the next budget decision

Write the decision, the business outcome and the counterfactual you need to understand. Then select the lightest credible method that can answer it, and record what remains unknown.