Three Ways to Measure the Same Sale

This comparison asks what attribution, marketing-mix modeling, and randomized incrementality each do with the same customer journey. The decision it supports is conditional: which method is appropriate depends on whether the brand needs visible touchpoint management, an aggregate response across channels, or a causal estimate for the next budget decision. The methods are not rival names for one number. They observe different outcomes, use different comparisons, and fail in different places when demand created by the generator is completed on the trusted marketplace.

The same journey, three measurement questions

The generator's ad can create intent, the customer can complete the purchase on the marketplace, and the brand can still receive several legitimate-looking reports. Attribution asks which observable advertising touchpoint should receive credit for a conversion the system can match. MMM asks how aggregate sales vary with channel spend over time or geography. Incrementality asks how outcomes differ between a randomized treatment assignment and a comparable control condition. The fact that all three reports may contain the word “return” does not make their numerators or causal meanings equivalent.

Attribution: who gets credit for the visible conversion?

Attribution is built around observability. A pixel, conversion API, click log, identity link, or similar signal connects a transaction to an ad touchpoint. A last-touch rule may assign the conversion to the final observed interaction. A multi-touch rule may distribute it among several observed interactions. A data-driven system may use more elaborate patterns. The common feature is not the particular crediting heuristic; it is that the conversion must enter the observable set before the heuristic can act on it.

That makes attribution useful for operational questions that are explicitly about visible response. A campaign manager can see which ads or placements are associated with recorded conversions, monitor changes quickly, and adjust creative or delivery within the system's reporting window. The metric can describe how the generator's observable customers behave. It can also reveal that the marketplace is present near many completed orders.

But attribution cannot answer the causal question merely by becoming more detailed. If a customer discovers a product through G and completes the purchase on R without a recoverable signal, the conversion is absent from G's attribution universe. Last-touch logic cannot credit it to G; multi-touch logic cannot divide it; a more complex model cannot observe it unless additional evidence or assumptions are introduced. The method can redistribute visible credit, not recover invisible demand.

This is a limitation of the numerator, not necessarily of the arithmetic. A high attributed ROAS means that the reported revenue is high relative to the spend under the matching rules in force. A low attributed ROAS means that the reported revenue is low under those rules. Neither statement alone establishes the number of purchases that would disappear if the advertising were removed. The generator's true contribution may be understated when its customers are more likely to cross into a marketplace checkout.

MMM: how does total sales move with channel spend?

A marketing-mix model changes the outcome boundary.