Conversion Modeling is Meta and Google's machine-learning layer that estimates conversions on traffic where deterministic attribution is unavailable, using observed behavior on consenting users to extrapolate to unattributed cohorts.
When a user sees your ad but doesn't click, or clicks from an iOS device that opted out of tracking, the conversion platform has a gap. Conversion modeling is how Meta and Google fill it: a machine-learning layer that estimates the conversions on traffic where deterministic, user-level attribution is unavailable.
The mechanism works by observing consenting users whose pixel and CAPI data flows cleanly. From that observable cohort, the platform builds a statistical model of conversion probability given ad exposure, device type, time of day, audience segment, and a dozen other signals. That model is then applied to the unattributed cohort — the iOS users, the cookieless browsers, the consent-declined sessions — to produce an estimated conversion count.
On Meta, this process became structurally mandatory after iOS 14.5 and the ATT prompt. SKAdNetwork postbacks provide aggregate counts but strip user-level detail; the platform falls back to its consenting-user model to extrapolate the gap. The quality of that extrapolation depends directly on your Conversion API (CAPI) coverage: the more server-side signal you send, the better the consenting cohort your model trains on. Google's data-driven attribution model uses the same pattern — observed multi-touch paths train an ML layer that estimates contribution for the cookie-less remainder.
By mid-2025, conversion modeling had expanded further. Meta's Andromeda delivery architecture, Advantage+ campaigns, and Google's Meridian MMM framework all assume a baseline of modeled data. We see accounts where 35–50% of reported conversions are estimated, not observed. That share rises as consent rates erode and iOS penetration grows. For a full picture of what this means for your reporting stack, the post on the death of attribution in 2026 covers the structural shift in detail. And if you're cross-referencing platform estimates with an external ground truth, AI analytics tools for marketing includes a breakdown of which tools handle modeled-conversion uncertainty well.
Treat modeled ROAS as a directional signal, not a financial figure — always anchor it to a lift test or MMM before acting on it.
A growing share of ad-attributed conversions are now modeled, not observed. Conversion modeling is the engineering reason your iOS reported ROAS no longer collapses to zero after an ATT opt-out wave. It is also why small calibration drifts — a CAPI coverage drop, a consent-rate change — can push reported numbers far from reality with no creative or audience change at all. If your ground-truth anchor (post-purchase survey, MMM, conversion lift test) diverges from platform numbers by more than 20%, the model has drifted.
