ROAS Shows Where a Sale Was Credited, Not What Caused It
The subject of this series is media measurement: how a brand decides which advertising activity created value and how much to spend on it next. The specific problem is the gap between a sale being credited to an observable ad touchpoint and an advertisement causing an incremental purchase. That gap matters because the number in a dashboard is not merely descriptive. When it controls the next budget, an incomplete measurement can change the market it is measuring. The practical judgment this series supports is simple to state but demanding to apply: before treating ROAS as evidence of effectiveness, ask what part of the customer journey the metric can actually see.
Imagine a customer encounters a product while scrolling through a short-video feed. The ad creates interest. The customer later opens a familiar marketplace, uses a saved payment method, and buys there because the marketplace feels safer or more convenient than the newer storefront attached to the feed. The purchase is real. The advertising may have been decisive. Yet the brand's attribution system may record the marketplace's sponsored placement—or no advertising touchpoint connected to the original discovery at all.
Nothing about this situation requires the recorded events to be fabricated. The marketplace really did host the checkout. The originating platform really did show the ad. The problem is that the customer journey crosses the boundary of the measurement system. A report that sees only the final, matchable touchpoint can be accurate about what it observed while remaining incomplete about what caused the purchase.
One purchase, several participants
The relationship is easiest to understand when the actors are kept distinct. A brand wants to sell a product and allocates advertising money. A demand-generating platform reaches people and may operate its own, relatively new storefront. A trusted retailer or marketplace offers another place to transact, often with familiar payment, shipping, returns, and buyer protection. The customer moves through these surfaces, but the measurement system does not necessarily follow them all.
The demand-generating platform and the marketplace therefore perform different economic jobs. The first may create or strengthen purchase intent. The second may make the final transaction comfortable enough to complete. In ordinary language, one can set up the purchase while the other closes it. A measurement system that assigns value only to the closing surface will tend to call the second job “performance” and leave the first job undercounted.
That is the central subject here—not a contest between two websites, and not a claim that every marketplace purchase was caused by an earlier ad. The useful question is conditional: when an advertisement creates incremental intent, what happens if trust causes the customer to complete the purchase elsewhere? The formal model behind this question treats the generator's ad spend as producing incremental purchase intent with diminishing returns. A separate trust parameter determines whether that intent ends in the generator's storefront or in the trusted marketplace. Total incremental purchases can stay the same even as their booking location changes.
This distinction also explains why a brand can face conflicting accounts of the same campaign. The generator's platform may report few conversions because it cannot observe the marketplace checkout. The marketplace may report strong returns because its sponsored placement appears near the completed sale. The brand's total transaction records may show that demand rose, but an aggregate sales model may still struggle to determine which channel deserves the response. Each account is built from a different boundary around the journey.
The boundary becomes consequential when the brand uses the report to decide where the next dollar goes. If the platform that generated intent looks weak, the brand may reduce its spend there. If the marketplace appears to harvest the resulting demand, it may receive more. The measurement system then does more than describe a journey after the fact: it feeds back into the conditions under which future demand will be generated.
The dashboard’s boundary
A dashboard often compresses three separate events into one row called a conversion. First, an advertisement may change a person's intention or behavior. Second, a channel may host and record the transaction. Third, a reporting system may match that transaction to an advertising touchpoint and assign credit. These events can coincide, but they do not have to.
The first event is about causation. It asks whether the purchase would have happened without the advertising intervention. The second is about commerce. It asks where the customer actually paid. The third is about observability. It asks whether the brand can connect the payment record to the earlier ad under the identifiers, windows, pixels, APIs, or other signals available to its reporting system.
ROAS—return on ad spend—is usually calculated from the revenue assigned to an ad divided by the spend on that ad. That definition makes ROAS useful for understanding the return represented inside a particular measurement boundary. It does not, by itself, say that the credited ad caused all of that revenue, nor does it guarantee that all revenue the ad caused was eligible for credit.
Suppose a customer buys on the generator's own storefront. In the model, that purchase is observable to the generator's attribution system. Suppose instead that distrust of the on-platform checkout sends the customer to the marketplace. Unless the marketplace passes a usable signal back, the generator's system may never see the connection. The purchase has not disappeared from reality; it has disappeared from the generator's attributable universe.
This is why changing the crediting rule is not enough. Last-touch attribution might give the final visible interaction all the credit. Multi-touch attribution might divide that visible credit among several observed interactions. A data-driven model might use more elaborate patterns. But none of these methods can redistribute a conversion that was never observed in the first place. They can change who receives credit for the visible remainder; they cannot recover an invisible path.
The causal and reporting questions can be placed side by side:
- Demand generation: Did the ad create an incremental reason to buy?
- Transaction booking: Where did the customer complete the purchase?
- Signal matching: Can the transaction be connected to the ad?
- Budget decision: What will the brand fund after seeing the report?
A single ROAS number may be a reasonable answer to the third question and a poor answer to the first. The danger begins when the answer to the third is used as if it settled the first and then determines the fourth.
The diagram shows why a measurement problem can become an allocation problem. The customer takes one path through the commercial system, but the brand's evidence follows only the branches it can match. The next budget decision is made from that partial record and can send the following cycle toward the branch that was easiest to observe rather than the one that created the most incremental demand.
Why the distinction changes the decision
The economic stakes do not depend on the brand choosing an absurd metric. Attribution-based ROAS is fast, familiar, and useful for managing visible responses. The problem is its silent substitution of observability for causality. If the generator's conversions are systematically less observable than the marketplace's, the generator can look worse even when it is creating valuable demand.
That creates a particularly damaging feedback loop. The generator creates intent; trust diverts some buyers to the marketplace; the reporting system misses part of the generator's contribution; the brand lowers the generator's measured ROAS; and the next allocation shifts away from demand generation toward the channel that records the transaction. Over time, the system can reward harvesting existing intent while starving the activity that creates new intent. A platform can be penalized not because its advertising failed, but because its success completed on another surface.
The same sale can therefore support three different kinds of statement. Attribution can say, “This touchpoint received credit for the revenue we matched.” A marketing-mix model can say, “Total sales moved with these channels over time and geography,” subject to the variation and endogeneity that make the channel response identifiable. A randomized incrementality study can ask, “How did outcomes differ when comparable audiences were assigned to receive the advertising opportunity rather than held out?” The statements are not interchangeable because their outcomes, comparisons, and blind spots differ.
There is a second boundary to keep in view. Even randomized measurement is only as strong as the outcome it measures and the report it releases. If the experiment counts purchases only on the originating platform, the off-platform assist can remain hidden inside the test. If privacy systems later remove matches, suppress small cells, or add noise, a clean causal design may yield a report that cannot support a fine-grained decision. A strong measurement practice must therefore preserve two distinctions: credit versus cause, and causal identification versus the amount of evidence still observable after reporting constraints.
The first conclusion is not that ROAS is useless. It is that ROAS should be read with its denominator, attribution window, outcome coverage, and matching boundary in view. A number can be directionally helpful for the records it includes and still be systematically misleading as a guide to incremental budget. The next question is consequently concrete: when intent travels from one surface to another, how large can the missing share become, and how does trust determine what the dashboard records? Understanding that journey is the step that turns a warning about metrics into a testable explanation of their behavior.