WebmasterID logoWebmasterID
Attribution models

Holdout-based attribution

Holdout-based attribution uses a randomized holdout — a group deliberately excluded from a campaign or channel — to estimate how much of a channel's credited conversions are genuinely incremental. By comparing the treated population against the holdout, it grounds attribution in a counterfactual rather than relying solely on observed click paths, which tend to over-credit channels that intercept already-converting users.

Partially verified

What this means

Touch-based attribution credits whatever interactions it observes, but it cannot see the counterfactual: would those conversions have happened without the channel? Holdout-based attribution adds that missing piece by randomly excluding a portion of users (or regions) from a channel and measuring the resulting change in conversions.

The measured incremental share can then be used to adjust the credit a rules- or data-driven model assigns. A channel that loses little when held out was largely harvesting demand; a channel whose holdout drops sharply was genuinely generating conversions.

Where it helps and its limits

Holdouts are especially valuable for channels prone to over-attribution — brand search, retargeting, and any tactic that re-engages users already on a path to convert. Combining holdout results with a touch model gives credit that is both granular (per touch) and grounded (calibrated to incrementality).

Limits: holdouts require giving up some media on purpose, enough scale to detect change, and a clean test window; they yield aggregate incremental shares, not per-impression truth. Treat holdout-based attribution as a calibration layer over a model, not a replacement for sound experimental design.

How it appears in analytics and logs

If holding out a channel barely changes conversions, much of its attributed credit was non-incremental; a large drop confirms the channel was genuinely driving conversions.

Diagnostic use case

Apply a holdout to a channel you suspect is harvesting existing demand (such as brand search), then scale its attributed credit toward the measured incremental share.

What WebmasterID can help detect

WebmasterID's first-party conversion events can provide the outcome series for treated and held-out populations used to recalibrate attributed credit.

Common mistakes

Privacy and accuracy notes

Holdouts compare aggregate group outcomes, not individual identities. This page is educational and not statistical or legal advice.

Related pages

Sources and verification notes

Last reviewed 2026-06-24. Facts are checked against primary/official sources where available; uncertain specifics are marked “Data not yet verified” rather than guessed.