Retargeting often reaches people who already have a reason to return.
They visited the store. They compared products. Some put an item in a basket. If they buy after seeing an ad, an attribution report may credit the campaign. That record alone cannot tell you how much the campaign changed their behavior.
You may be paying a doorman to open a door people were already walking through.
That possibility deserves a test. It does not justify switching off a channel on instinct.
A well-designed holdout assigns eligible people to treatment and control before exposure. One group is eligible to receive the tested advertising; the other is held back. The comparison estimates the additional outcomes associated with that advertising, subject to the experiment's design, uncertainty, and limitations.
The groups will not be identical. Random assignment helps make them comparable in expectation. Sample size, contamination between groups, and the time needed for purchases to emerge still matter. Where person-level assignment is unsuitable, a carefully designed geographic experiment may offer another route.
There is an equally important limit: a group-level experiment does not reveal the unobserved alternative history of each individual buyer. It cannot stamp a particular order with certainty that says, “This person would never have bought without the ad.”
A prediction that someone was likely to buy does not settle that question either. Purchase probability and the effect of advertising answer different questions. Someone who is already interested can still respond to an intervention.
This is why the causal record needs to keep the estimate, its uncertainty, and its assumptions together.
MIZ OKI's Causal Credit Ledger is intended to preserve that distinction between platform attribution and evidence of incremental effect. The causal core is currently labeled partial. Ghost bids remain proposed. Those labels describe capability maturity, not the strength of any particular experiment.
In the Media demo, Act 3 illustrates the sequence: register the holdout, inspect the estimate, then challenge the conclusion. The examples are synthetic. Refutation checks can expose fragile assumptions; passing them does not, by itself, establish causality.
The budget question comes next.
Even a positive estimate of incremental revenue does not establish that the campaign is profitable. Returns, product costs, fulfillment, and the cost of the advertising still have to be paid. A revenue-based incremental ROAS above one can coexist with a loss.
Nor does a weak result automatically mean the right budget is zero. The evidence may support reducing spend, redesigning the audience, running a better-powered experiment, or holding until uncertainty narrows. The alternative use of the money matters too.
Ad platforms also offer lift studies. The useful distinction is between ordinary attribution reporting and an experiment that answers a causal question. MIZ OKI's purpose is to carry qualified evidence into the wider business decision, with economics and authority attached.
The doorman metaphor is a question about marginal value. What did this spend add, what did that addition cost, and what evidence would change our decision?
Open Act 3 of the Media demo and examine the example that withholds a conclusion. For the product's causal and intent principles, visit Signal.