What are modeled conversions, and how should you treat them?
Modeled conversions are conversions a platform estimates statistically rather than observes, filling the gap left by consent refusals, cross-device journeys and browser restrictions. Estimating is usually the honest response to a genuinely unobservable question. The problem is presentational: the estimates arrive in the same column as directly measured conversions, with nothing marking which rows are which.
Google labels part of its reported conversions as modeled, and Meta's aggregated event measurement performs a similar reconstruction for iOS traffic where users declined tracking. Neither platform is inventing demand. They are answering a question that stopped being directly answerable when consent frameworks and browser restrictions removed the underlying observations, and a good estimate is more useful than a total that silently excludes everyone who declined.
What follows from that is a discipline about precision rather than a reason for distrust. A conversion total that mixes observed and estimated rows carries uncertainty that is invisible in the number itself, so quoting it to the pound implies a confidence nobody has. Small differences between two modelled figures, especially across a segment breakdown, are frequently smaller than the modelling error and should not carry a decision on their own. Breakdowns deserve more suspicion than totals, because models are fitted to reproduce aggregate behaviour, not every slice of it.
The reliable move is to hold one number the business genuinely owns. Whatever the platforms report, the payment processor, booking system or CRM records what actually happened, and reconciling channel totals against that on a regular cadence tells you the size of the gap you are working with. A stable, known gap is workable. An unknown one means every downstream figure inherits it.