Why holdout
Attribution models estimate. Holdouts measure. The randomisation eliminates the selection bias that makes pixel-based attribution unreliable on iOS, in incognito, and after consent rejection.
What the platforms support
Meta supports Conversion Lift studies natively. Google supports geo and audience holdouts via Experiments. Both expose the random assignment via API; Ripplux reads the result and updates the ROAS engine.
What it costs
The control arm doesn't see ads, so the merchant gives up some conversions during the test window. The trade-off is a clean number that future budget decisions can rely on.