Attribution vs incrementality: two different questions

Start with the decision you need to make

If the question is which source context belongs to an inquiry, attribution is relevant. If the question is whether the business would lose sales without a channel, you need evidence about the alternative outcome. Those are different measurement jobs.

An ad can be part of a customer’s path even when they already intended to buy. A referral can also create demand that later appears as a paid-search conversion. A useful report should explain what its model measures rather than label every credited sale as additional growth.

Compare the methods by the question they answer

None of these methods removes the need for sound outcome definitions. If Won means one thing in one month and another thing in the next, a sophisticated model can still be built on an inconsistent target.

Match the decision to the strength of the evidence

Suppose a branded campaign receives credit for many sales. That can help explain the recorded purchase path, but it does not establish how many customers would have purchased without that campaign. Treating all credited revenue as additional revenue can overstate the campaign’s causal contribution.

An incremental-impact question needs a suitable comparison and a plan for interpreting it. Depending on the context, that may involve a controlled holdout or another defensible causal method. The design needs to account for the outcome, timing, sample limitations and possible spillover. A simple before-and-after change can mix the intervention with seasonality or other changes.

For day-to-day operations, attribution can still reveal broken handoffs and useful source context. Use it for the questions it can support, and label stronger causal claims separately. If the business cannot run a credible experiment yet, record the uncertainty and make a bounded decision rather than presenting a dashboard model as proof of lift.

Before calling a result incremental

• State the intervention, target population and outcome being evaluated.
• Explain the comparison and why it represents a credible counterfactual.
• Record concurrent changes, timing and major limitations.
• Keep observed, attributed and estimated incremental outcomes clearly labelled.

Why a before-and-after chart is not enough

Sales can change because of seasonality, pricing, availability, competitors or sales execution. Comparing a month with advertising to a month without it does not automatically separate those influences.

Google’s Meridian GeoX documentation describes estimating the counterfactual: what would have happened to the treated group without the intervention. The observed result alone cannot supply both sides of that comparison. A valid design needs an appropriate way to estimate the alternative.

Decide whether an experiment can answer the question

1. Name one decision and the outcome that would change it.
2. Define the intervention, comparison and observation period.
3. Assess whether the design has enough information to detect a useful effect.
4. Identify contamination, seasonality and operational changes that could undermine interpretation.
5. Predefine the analysis and how uncertainty will affect the decision.

A small service business may not have enough volume or geographic separation for every experiment. That is a reason to assess the design, not to pretend a noisy result is decisive. Specialist measurement work can be appropriate when the budget decision justifies it.

Use operational evidence without overstating it

CRM source context, qualification and recorded outcomes are still valuable. They can reveal duplicate events, poor-fit inquiries and missed sales work. Those findings support operational improvements even when causal advertising lift has not been measured.

Growth Control implements this connected operating process. Its current website does not claim that every engagement includes a proprietary multi-touch model, marketing mix model or incrementality-testing engine. Where a causal decision requires more, the method and scope need to be defined explicitly.

Questions and answers