Where User-Level Incrementality Stops Working

The subject here is not whether an advertiser can build a more elaborate lift dashboard. It is whether a brand can assign the advertising opportunity to a known unit before the generator's platform decides who sees it. That boundary determines when user-level ITT is meaningful, why two tempting after-the-fact fixes fail, and when geography or time can provide a coarser but still valid comparison. The practical judgment is a feasibility check: establish the assignment universe and outcome scope before trusting an individual-level result or projecting it from retargeting into acquisition.

The assignment must exist before the exposure

For user-level randomization, the advertiser needs an enumerable assignment universe at the moment an audience activates. “Enumerable” does not mean that the brand must know a person's name and postal address. It means that each potential unit has an individual or pseudonymous identifier that can be placed into a treatment or control bucket before the platform delivers the advertising opportunity. A mobile advertising identifier or, with limitations, a first-party cookie can serve this role when it is addressable on the platform.

That timing is the foundation of the contrast. The brand assigns a known audience member to treatment or control, and then the generator's delivery system operates within that assignment. Some assigned users may see the ad and some may not. The ITT asks about the effect of being assigned the opportunity, not the effect among the people the platform later selected for exposure. If the brand conditions on exposure, it conditions on a decision made by a system that may already be using predicted purchase behavior, bids, engagement, or other signals related to the outcome.

The distinction matters in the familiar generator-and-marketplace journey. Suppose the brand assigns part of an enumerable audience to receive G's advertising opportunity. Some treated customers discover the product in G and complete the purchase on R because they trust R's checkout. If the brand's outcome includes both destinations, the assigned treatment group retains those purchases. The causal comparison is still available even though the generator's own attribution feed may not show them.

If the brand instead compares only users whom G actually exposed, it may be comparing a selected group of likely recipients with a control group that was not selected under the same rule. The selection is no longer controlled solely by the experiment. A high propensity to receive the ad can be related to a high propensity to buy, and the resulting difference can look like advertising lift even when it partly reflects delivery selection. Assignment before exposure prevents that particular conditioning error by defining the comparison before the platform's delivery process runs.

Two common repairs fail for opposite reasons. The first is post-hoc assignment. A brand observes who converted and then randomly labels those converters as treatment or control. Those labels are independent of the earlier advertising opportunity by construction. They cannot tell the brand whether the ad changed the conversion probability, and treating them as an ITT group dilutes the real contrast toward zero.