How modelled conversions work
An advertising platform can only count what it observes. A click it can attribute, a tag that fires, an identifier it can read again later — those produce an observed conversion. When part of that chain is missing, the observed count falls short of what really happened: someone declines analytics consent, a browser blocks the identifier, or a person clicks on a phone and buys on a laptop.
Modelling closes that gap with inference rather than invention. The platform studies the journeys it can still follow end to end, learns how that traffic behaves, and applies the same pattern to comparable traffic it cannot follow. The estimate is then added into the same conversion column as the observed figures, usually without a separate label, which is why many advertisers never realise part of their reporting is calculated rather than counted.
Why modelled conversions matter
Automated bidding optimises towards whatever the account reports. If measurement loss quietly removes conversions, the system reads that as weak demand and bids down on traffic that is in fact converting. Modelling exists largely to stop that spiral, which is why it shows up most in the accounts that lean hardest on automated bidding.
It also changes how a report should be read. Two channels are no longer being measured the same way once one of them is filling gaps and the other is not, so a straight comparison of conversion counts inside an analytics and tracking setup can flatter whichever platform models most willingly.
Where modelled conversions go wrong
The first mistake is treating a modelled conversion as a person. You cannot export it, call it, or find it in your CRM, because it stands for a share of unobserved activity rather than one identified customer. Reconciling platform totals against a sales list line by line will always fail, and that failure is not a fault in the setup.
The second is adding platforms together. If several accounts each model the same partly hidden journeys, the combined total counts the same sale more than once. The third is assuming modelling always applies. Models need enough observed data to learn from, so a small account in a thin market — which describes plenty of advertisers in Nepal — may get very little modelling and see a blunt drop in reported conversions instead.
What to do about it
Fix observation before arguing with the estimate. A correct consent mode implementation, conversions sent server-side or by API, and clean first-party identifiers all give the model better material to work from and shrink the part that has to be inferred.
Then judge the account at the business level. Keep enquiries and sales in your own records, compare the direction of platform reporting against them, and use the platform figures for deciding where budget moves rather than for reporting revenue to anyone. When a gap opens up, work out whether it is a genuine data discrepancy in the setup or simply two systems counting different things by design.
