How modeled conversions work
When a user does not consent to tracking, converts on a different device, or uses a browser that blocks the identifier, Google can no longer draw a straight line from ad click to conversion. Instead it uses aggregated, observable behaviour from users who could be measured to estimate how many of the unmeasured interactions converted. That estimate is a modeled conversion. The output is blended into your normal conversion columns, which is why a reported total is often part observed and part modeled without an obvious label.
Why they exist now
Modeling grew directly out of privacy changes. Consent banners, the decline of third-party cookies, and browser restrictions on cross-site identifiers all remove signal that used to be deterministic. Rather than simply under-reporting, platforms model the missing conversions so bidding algorithms still receive a complete-enough picture to optimise. The upside is that Smart Bidding is not starved of data. The trade-off is that a modeled number is a probabilistic estimate, so it will not tie out to a source system row for row the way an observed conversion does.
When to trust them
Modeled conversions are most reliable when the model has plenty of observed data to learn from, which means high-volume accounts with strong first-party signal. They are least reliable on thin-volume accounts or where consent rates are very low, because the model is extrapolating from a small observed base. The practical move is to know your modeling rate, keep it stable, and reconcile the observed portion against a source of truth like GA4 or the backend, rather than trusting or dismissing the whole number.
How iClick treats modeled conversions
iClick treats modeled conversions as useful for bidding but never as the sole source of truth for reporting. The approach is to strengthen the observed signal first, with Consent Mode and Enhanced Conversions feeding better first-party data so the model has more to work from, then reconcile the deterministic portion against backend revenue. The rule is to reduce how much has to be modeled before trusting the model, because a smaller estimation gap is an easier one to trust.
