Attribution and Incrementality

Every advertising platform will report the sales it was involved in. Every platform will also claim credit generously, because the alternative is visibly worse numbers. The result is that reported attribution is a description of how a platform chooses to assign credit, not a measurement of what your advertising did.

The two ideas you need

Attribution is the question of which advert gets the credit for a sale. Every platform answers it with its own model, and the models disagree. There is no correct answer available from inside the data.

Incrementality is a different and more useful question: what would have happened without this spend? This is not what attribution is measuring, and it is usually the question a business actually cares about.

Why attribution and incrementality diverge

Some sales would have happened regardless. A person who was already going to buy the book this week may have seen an advert, and the platform counts it. Some sales genuinely were caused. Neither platform can distinguish these from inside its own data, because it cannot see what the person did in the counterfactual case where they saw nothing.

What you see instead is a set of models that differ by rule rather than by truth, and the rules are tuned so each platform's number looks good in its own reporting.

The experiment that settles it

Hold out a portion of audience or budget. Run the campaign as normal, and alongside it withhold spend from a randomly selected slice of the eligible audience. Compare total sales between the two groups. The difference is the incremental effect, and it is the only number here that means what you want it to mean.

This is a real experiment with real cost, which is why it is rarely done and why most reported returns are optimistic. Where it has been run, the gap is frequently substantial.

How to use reported attribution honestly

Use it to compare campaigns within the same platform, with the same model, over similar periods. Use it to spot where attention is going. Do not use it to compute whether a channel is profitable, and do not sum it across platforms, because you will be adding up several incompatible rules.