A Better Measurement Contract
The opening question was whether a sale credited to one surface tells the brand what caused it. The answer is no—not because credit is always worthless, but because the generator platform, trusted marketplace, and brand budget sit on different parts of the customer journey. A better measurement contract connects the claim to the outcome, comparison, reporting quality, and decision threshold that can support it. It also states what to do when the evidence is not strong enough: improve the measurement, narrow the claim, or leave the decision unresolved rather than manufacturing certainty.
Keep every claim attached to its evidence
The most useful discipline is to label a result before interpreting it. Consider a customer who discovers a product on the generator, completes the purchase on the marketplace, and appears in several systems. The label attached to the record determines what the brand is entitled to say.
A credited sale is a transaction matched to an advertising touchpoint under a particular attribution rule and reporting window. It can tell the brand which visible interaction received credit. It cannot, on its own, say how many purchases would have occurred without the ad. If the marketplace does not return the signal from the checkout, the generator's missing record is not evidence that the ad created no demand.
An aggregate response is a relationship between total sales and channel spending over time or geography. It can retain purchases booked on the marketplace, so it may see more of the journey than attribution. But its channel interpretation depends on independent variation and a model that can separate the generator's demand from the marketplace's response to arriving intent. An aggregate coefficient is not automatically an individual effect or a causal ownership label.
An experiment-identified lift is a treatment-control contrast based on randomized assignment and a defined outcome. When the outcome is channel-complete, it can answer whether the advertising opportunity changed total purchases even if trust sent treated customers to the marketplace. The label identifies the causal contrast under the design's population, assignment, outcome window, and interference assumptions. It does not automatically identify every audience, creative, or future market.
A predicted incrementality score extends experiment-identified evidence to a user, audience, or campaign that was not itself held out. It can help rank users, allocate scarce experimental attention, or project a response beyond direct coverage. It remains a prediction conditional on feature support and a stable feature-to-effect mapping. A score is not a new randomized result for the person who receives it.
A privacy-certified effect is a threshold claim that survives the declared ambiguity about signal loss and the finite-sample uncertainty in the released report. It is stronger than a positive point estimate because every compatible clean world still clears the business threshold. A privacy-unresolved effect is not a null. It says that the report is compatible with clean worlds on both sides of the threshold, so a uniformly valid procedure cannot choose a stronger claim from that report alone.
These labels are not bureaucratic additions to a dashboard. They prevent the brand from transferring authority from one measurement regime to another without noticing. A generator's attributed ROAS may be useful for visible delivery management but too weak to justify reducing a demand-generation budget. A marketplace's order record may be accurate about the checkout but too weak to prove that its sponsored placement created the purchase. A model score may be useful for choosing the next experiment but too weak to authorize an untested acquisition population. A certificate may support action at one reporting grain while leaving a subgroup claim unresolved.