Attribution counts who was touched. Incrementality counts what was caused.
Same data. Three slicings. A three-fold gap.
A holdout experiment shows 100 conversions with marketing off and 110 with marketing on. Three ways to slice that observed lift produce three different "marketing impact" numbers. Of the 30 conversions credited by attribution, 20 would have happened anyway and 10 are caused by marketing — that's the bridge between the two views.
The figure · same 110 conversions, three slicings
Both views are looking at the same 30 touched conversions. The difference is what they count.
Attribution credits all 30. Incrementality only credits the part that wouldn't have happened anyway.
The clean 30 = 20 + 10 split holds only if marketing moved the people it touched and no
one else — no spillover to the untouched 80, a stable base rate; that's exactly what
Pollution and Substitution put under stress.
What counts as a touch — click only, view-through impression, 30-day window, cross-device — can swing the attribution number by 3–10×. The same channel reported under a 1-day-click rule and a 30-day-view rule gives wildly different "marketing impact." Attribution's 30 here is a stand-in for whatever your platform's rules count; the platform never measures the 10.
The 3× ratio shown here is illustrative. Real ratios vary widely with channel and audience: roughly 1.5–2× for cold-acquisition paid social, 3–5× for typical retargeting display, 10–20× for high-funnel view-through. The cleaner the audience (more net-new, less retargeted), the closer attribution and incrementality converge. When they diverge sharply, you're paying for conversions that were already coming. And the incremental 10 isn't read off a dashboard — it comes from a holdout, and like any experiment it carries a confidence band whose width is a question of how the test was sized.