When the platform 'models' your missing conversions, what are you actually buying?

When the platform 'models' your missing conversions, what are you actually buying?

The question: privacy changes (consent loss, app-tracking restrictions, cookie deprecation) mean platforms can no longer observe many conversions directly. Their fix: modeled conversions — statistically estimated stand-ins for the ones they can't see. Is this measurement, or sophisticated guessing you can't audit?

What's under the hood: modeled conversions typically use the observable conversions (from consented users) to train a model that predicts the unobservable ones, then reports a blended figure. Approaches range from simple uplift ratios to aggregated, differentially-private reporting and conversion-modeling pipelines. The methodology is legitimate — it's standard missing-data imputation. The catch is that you're now optimizing against predictions, and the prediction's assumptions are the platform's, not yours.

What to watch for: modeled conversions assume the unobserved users behave like the observed ones. When consented and non-consented populations differ systematically — different devices, regions, ad-blocking propensity — that assumption breaks, and the modeled total carries an unmeasurable bias. There's no error bar on most platform reports, which is itself a tell.

The nuance: this isn't a reason to ignore the numbers — un-modeled reporting now undercounts badly. The reason for rigor is that modeling shifts you from measuring to inferring, and inference needs validation.

What to actually do: treat modeled conversions as estimates with hidden variance. Anchor them to incrementality experiments and MMM, which don't depend on individual-level tracking, periodically.

Bottom line for practitioners: modeled conversions are imputations, not observations. Use them — but validate them against tracking-independent methods, because you can't see the error you're optimizing into.