The attribution problem no one talks about

By Mo Touzani · 2025-05-14 · Marketing Measurement

Every company with a multi-channel marketing budget has an attribution model. Most have three or four, because none of the models agree with each other and no one wants to admit that the disagreement is fundamental rather than technical. The real attribution problem is not that the data is imperfect. It is that every major attribution model is answering the wrong question.

What attribution models measure

Last-click attribution gives full credit to the final touchpoint before conversion. First-click gives it to the first. Linear spreads credit evenly across all touchpoints. Position-based weights the first and last more heavily. Data-driven uses machine learning to weight touchpoints based on historical patterns. All of these models answer the same question: which channels were present in the conversion path? None of them answer the question that matters: which channels caused the conversion?

Presence and causation are different things. The path a buyer took to convert includes every touchpoint they encountered. It does not tell you which of those touchpoints would have changed the outcome if it had not existed. A buyer who was going to purchase regardless, who happened to click a retargeting ad in the final hour, looks identical in the attribution data to a buyer who was persuaded by that ad. Last-click attribution credits both conversions to retargeting. The model has no mechanism to distinguish between them.

This is not a technical limitation that better data would solve. It is a structural limitation of observational attribution. Without knowing what would have happened in the absence of each touchpoint, there is no basis for assigning causal credit. The models are not measuring the wrong thing because they are poorly designed. They are measuring the wrong thing because the right thing cannot be measured without a controlled experiment.

The counterfactual problem

To know whether a channel caused a conversion, you need to know what would have happened without it. This is the counterfactual. It cannot be observed directly. It can only be estimated through controlled experiments where the channel is withheld from a randomly assigned group and the difference in conversion rates is measured against a group that received it.

Attribution models do not run this experiment. They observe the paths that converted buyers took and construct rules or algorithms for distributing credit across those paths. The result is a ranking of channels by their presence in successful conversion paths. This is categorically not the same as a ranking of channels by their causal contribution to conversion.

The difference has direct financial consequences. Retargeting consistently receives high attribution credit because it operates at the bottom of the funnel, close to the conversion event. Incrementality studies consistently show that retargeting is heavily over-attributed, because a large fraction of the conversions it claims would have occurred without it. The buyers it claims to have converted were already in the market. The channel collected the credit but did not drive the decision.

Why the models disagree and what that tells you

When a company runs last-click, linear, and data-driven attribution simultaneously and the three models produce materially different channel rankings, that disagreement is typically treated as a measurement problem to be resolved. Which model is correct? This is the wrong question.

The disagreement between attribution models is not a measurement problem with a technical solution. It is a symptom of the fundamental limitation of observational data: without knowing what would have happened in the absence of each channel, no model can assign causal credit correctly.

The models disagree because they encode different assumptions about how credit should be assigned to touchpoints. Last-click encodes the assumption that the final touchpoint is most important. Linear encodes equal importance. Data-driven encodes the assumption that historical conversion patterns reveal the true weights. None of these assumptions has been validated against a controlled experiment. None of them is demonstrably correct.

What the budget conversation should look like

Attribution data is useful for understanding purchase path patterns, identifying frequency and reach issues, and diagnosing where buyers are dropping out of the consideration set. It is not useful for ranking channels by causal contribution to revenue, which is what most attribution conversations are used for.

The appropriate tool for causal channel ranking is incrementality testing: controlled experiments where channels are withheld from randomized groups and the revenue difference is measured. This is harder to execute than pulling an attribution report. It is the only method that answers the question the attribution report is being used to answer.

Most companies run attribution reports. Very few run incrementality tests. The result is a systematic misallocation of budget toward channels that are most visible in the conversion path rather than most responsible for causing conversions. Companies that close this gap have a structural informational advantage over those that do not.

The framework for honest measurement

1. Stop asking which channel gets the credit. Start asking which channel changes the outcome. Credit allocation is an accounting exercise. Causation is a strategic question. Reframe every attribution conversation around counterfactuals: what would conversion rates look like if this channel did not exist? If you cannot answer that question from your current data, you need a different measurement approach, not a better attribution model.
2. Use attribution data for what it is good at. Attribution models are strong at revealing reach and frequency patterns, touchpoint sequencing, and funnel drop-off points. Use them for those purposes. Do not use them to rank channels by causal contribution or to justify budget shifts between channels. Those decisions require experimental evidence.
3. Build an incrementality testing calendar before the next budget cycle. Incrementality tests require planning time, minimum traffic volumes, and organizational alignment on what the test is designed to measure. Design the tests before the pressure of a budget decision forces you to act on attribution data that cannot support the conclusion you need it to support.

What companies with honest measurement know

Companies that have invested in incrementality testing consistently find that their highest-attributed channels are not their highest-incrementality channels. The gap is usually largest for retargeting, brand keyword bidding, and direct traffic: all of which receive attribution credit for buyers who were already committed to converting.

Closing that gap does not require abandoning attribution models. It requires using them correctly: as diagnostic tools for path patterns, not as authoritative answers to causal questions. The companies that make this distinction allocate budget based on what drives incremental revenue, not what appears prominently in a conversion path report.