Modeled conversions

When a platform cannot observe a conversion — because the visitor refused consent, the browser blocked the mechanism, or the journey crossed a boundary the platform cannot see — it can either report nothing or estimate what it missed. Most now estimate, using the observable conversions as training data.

The reasoning is defensible. If a measurable slice of a comparable audience converts at a certain rate, the unmeasurable slice probably behaves similarly, and reporting zero for them is also wrong — arguably more wrong. A platform that reported only what it could see would systematically understate itself.

What this means for your numbers. Your dashboard total is partly counted and partly inferred, and the split is usually not shown. That has three consequences worth internalising:

• The total will not reconcile with your own database, and chasing the difference is often a search for something that was never a discrete event.
• The model can be updated retroactively. Numbers you already reported can move without anything having happened, which is disorienting the first time and awkward if you put them in a board deck.
• You cannot audit it. The methodology is the platform’s and is not published in a form you can check.

The thing to hold on to: modeled conversions are an estimate of what was observed but missed. They are not a measure of whether the advertising worked — that is incrementality, and no amount of modelling by the seller answers it. A platform modelling its own contribution has an obvious direction to be wrong in.