How modelled data works
Measurement platforms no longer see every visit. Cookie limits in browsers, consent choices, app tracking rules and people moving between a phone and a laptop all leave holes in what can be observed directly. Rather than report those holes as zero, Google, Meta and other platforms fill them with estimates built from the traffic they can still see.
The mechanism is pattern matching. The platform takes the behaviour it does observe, works out the relationship between that observed group and what actually happened, then applies the same relationship to the traffic it could not follow. The output is added into reports beside observed figures, often without a label separating the two.
Which parts of a report are modelled varies by platform and by report. It can affect conversions, user counts and the way credit is split between channels. Some interfaces mark it with a note; some only mention it in the documentation.
Why modelled data matters
Without modelling, reports would understate results badly, and budget decisions would follow a measurement failure rather than the business. Filling the gap is an honest attempt to correct that. It is also the main reason platform totals rarely match a website’s own database, because each system estimates differently, or does not estimate at all.
It matters most when you compare one period with another. If the estimation method changes between them, the shape of your history changes with it, and that is a different thing from performance changing.
Common mistakes with modelled data
The first mistake is treating an estimate as a count. A modelled conversion is a considered guess about a person, not a receipt, and it cannot be matched line by line to an order in your accounts. The second is insisting platform and business figures should agree. They measure different things by different means, and they never will.
The third is starving the model. Weak consent handling, broken tags or a missing data layer remove the observed signal the estimate learns from, so the modelled share grows at exactly the moment it becomes least reliable.
How to act on it
Decide, before anyone builds a report, which dataset settles which question. Money spent and orders delivered belong to your own records. Direction, channel comparison and day-to-day optimisation belong to the platform, estimates included. Write that split down so one question is not answered two ways in two meetings.
Then protect the observed signal the estimates depend on: correct consent configuration, complete and consistent campaign tagging, and server-side measurement where you can implement it. Better first-party measurement makes the modelled portion smaller and the whole report steadier. That groundwork is what an analytics and tracking setup exists to fix, and the same reasoning applies to the modelled conversions you see inside an ad account.