Attribution vs. Incrementality: The Assisted Own Goal Problem

This distinction matters whenever a person discovers a product in one place and completes the transaction somewhere else. Short-form video may create demand. Branded search may capture it. A marketplace may close the transaction.

Last-touch reporting can reward the final measurable interaction while the demand-generating channel appears less effective than it really was. A July 2026 preprint describes one version of this failure as an assisted own goal: advertising causes an incremental purchase, but the conversion is booked on a downstream platform. The paper is a new methodological proposal—not settled industry consensus—but the decision problem is practical.

Attribution and incrementality answer different questions

Attribution is useful for fast operational monitoring. It gives teams granular campaign, creative and user-path signals. But attributed conversions do not establish a counterfactual. They do not tell you what would have happened if the campaign had not run.

Incrementality attempts to estimate that missing outcome. A randomized or quasi-experimental design compares observed results with a credible estimate of the outcome without the intervention.

A simple assisted-own-goal example

Depending on the rules, the app store, branded search or an MMP touchpoint may receive the conversion. The video campaign may receive partial credit or none. If the team pauses video because its attributed ROAS is below target, branded search efficiency may deteriorate later because less demand is being created upstream.

More attribution detail does not solve causality

Multi-touch attribution can improve path visibility, but it still assigns weights to observed interactions. Those paths are affected by targeting, prior intent, platform optimization and missing data. Users exposed to a campaign were usually not selected at random.

• High-intent users are easier to target.
• They are also more likely to convert without the ad.
• A campaign can therefore look effective while capturing demand that already existed.

MMM does not automatically identify the missing channel

Marketing mix modeling can estimate relationships between spend, media pressure and outcomes over time. It is useful for strategic allocation and saturation analysis. But causal credit still depends on variation, controls, model structure, priors and identification assumptions.

A decision hierarchy for budget changes

1. 01

   Validate the measurement chain

   Check links, SDK and consent changes, conversion windows, missing revenue events and web-to-app breaks.

2. 02

   Separate attribution from business outcomes

   Compare platform and MMP results with backend purchases, subscriptions, margin or retained revenue.

3. 03

   Look for causal evidence

   Use the strongest feasible design: lift test, randomized holdout, geo experiment or defensible natural experiment.

4. 04

   Calibrate the operating model

   Use periodic experiments to constrain daily planning and models without freezing one factor forever.

Five questions before cutting a channel

1. Did the business outcome fall, or only the attributed outcome?
2. Did another channel, marketplace or branded search capture more of the same demand?
3. Did tracking, consent, SDK or conversion routing change?
4. What experiment or natural variation supports the causal conclusion?
5. What result would cause the team to reverse the budget decision?

If these questions are unresolved, the correct state is not “the channel failed.” It is decision pending: measurement gap.

The operating principle

Use attribution for operational visibility, experiments for causal calibration, and MMM or response models for strategic allocation under explicit assumptions. The objective is not one perfect ROAS number. It is a budget decision that remains defensible when the lenses disagree.