The input metric determines the report's meaning — and how it should be used.
iaGMV. iGMV. Same data. A different question.
A two-cell holdout vs BAU test — a randomly held-out group that gets no marketing (the
holdout) versus business-as-usual (BAU) — read through two different
lenses. The lens is the metric you feed into the cell-difference calculation:
iGMV = E[Total GMV | marketing ON] − E[Total GMV | marketing OFF] — an
intent-to-treat lift on total revenue.
iaGMV = E[Attributed GMV | ON] − E[Attributed GMV | OFF] — the lift in your
attribution scorecard.
Both come from the same holdout. They measure different things and they answer different questions.
Each lift divided by spend is an iROAS (incremental Return On Ad Spend) — so the same
test yields an attributed iROAS and a topline iROAS that need not agree.
Attributed GMV
Attributed GMV
Topline GMV
In a clean holdout, the holdout cell has zero attribution events (no marketing → nothing
to attribute), so iaGMV ≈ BAU attributed GMV. The "i" prefix borrows
incrementality's credibility, but the math reduces to attribution. iaGMV inherits every bias of your
attribution rules — touch definition, lookback window, MTA weights. iGMV does not.
Run both lenses on the same test. The ratio is how much the attribution scorecard overcredits the channel — a single number that tells you whether to trust your existing dashboards.
Attribution lands within about 2× of the true causal effect. Trust the scorecards, with a small haircut as the ratio climbs. Typical for high-causal-share channels (cold acquisition).
Most attributed conversions are counterfactual. Don't use iaGMV for budget. Common for retargeting and view-through display (often 3–20×).
Channel is causing un-attributed conversions (rare). Attribution is under-selling impact. Often seen with upper-funnel brand work.
The practitioner rule: one lens per decision. Budget allocation, channel deprecation, ROI to finance — always iGMV from a holdout. Channel credit reports, journey analysis, MTA model calibration — iaGMV is fine because the question is about attribution mechanics, not causal revenue. When iaGMV and iGMV disagree by more than ~2× — the boundary above — your scorecard and your finance plan will tell different stories, and the finance plan is reading the right number for the budget call (provided the holdout itself is clean). One caveat the other way: iGMV is the unbiased lens but the noisier one — a topline holdout needs real sample size to resolve a small lift — while iaGMV is biased but precise.