What is Modeled conversions?

Modeled conversions are conversions a platform adds to its reporting because it believes they occurred, without having observed them.

They exist because the observable share of real conversions has fallen. Consent refusals remove some visitors entirely. Browser tracking prevention shortens or removes the identifiers that connect an ad interaction to a later purchase. In-app browsing contexts break continuity. Cross-device journeys — a phone click and a laptop purchase — cannot be joined without a signed-in identity.

In each of those cases a real conversion happened and the platform cannot see the connection. Reporting only what was observed would understate performance, misallocate budget toward whatever happens to be observable, and starve automated bidding of examples.

So the platform estimates: from the conversions it can observe, and the characteristics of the traffic it cannot, it infers how many conversions the unobserved portion produced.

An observed conversion corresponds to an event. A modelled conversion corresponds to a quantity.

That difference has practical consequences that catch accounts out.

It cannot be traced. There is no order to look up, no customer to check, no timestamp beyond the modelled attribution. Asking which order a modelled conversion refers to is a question with no answer.

It cannot be reconciled line by line. Aggregate comparison against the business's own records is the only route, and it is approximate.

It can move without delivery moving. When a platform updates its modelling, reported conversions change. The advertising did not.

It is produced by the measured party. The platform reporting its own performance is also estimating the part nobody can check. That is not an accusation of bad faith; it is a structural feature worth stating.

Find out the share. Where a platform discloses what proportion of reported conversions are modelled, that figure is worth knowing and worth watching. A campaign whose results are largely modelled is being reported on differently from one whose results are largely observed, and comparing the two as though they were the same measurement is a mistake.

Reconcile in aggregate. Total reported conversions against total orders for a period, monthly. The gap will not be zero. A stable gap is fine; a changing one is a signal.

Watch for steps. A conversion count that jumps without a corresponding delivery change is a methodology update or an implementation change, not a result. Noting the date prevents somebody drawing a trend line through it.

Keep the blended control. Total revenue over total spend starts from the bank and contains no estimates. When reported conversions rise and the blended ratio does not move, the modelling changed.

Do not stack estimates. Comparing one platform's modelled figures against another's is comparing two sets of assumptions, neither inspectable.

The share of conversions that has to be modelled falls as the share that can be observed rises, and that is largely within the advertiser's control.

Server-side event delivery gets events to the platform when the browser could not.

Hashed first-party identifiers attached to those events let them be matched to a known account rather than estimated.

Capturing the click identifier at arrival and storing it with the order gives the strongest available join.

Signed-in experiences, where the business can reasonably offer one, resolve cross-device journeys directly.

Each of these converts an estimate into an observation. The reported total may not rise much — the modelling was already filling part of the gap — and the reliability of the figure improves, and so does the quality of what automated bidding learns from.

Modelled conversions are a reasonable response to a real problem. The alternative, reporting only the observable, would be more auditable and less accurate, and would bias every decision toward the channels and contexts that happen to remain measurable.

They are also estimates, produced by an interested party, using methods that are not public and that change. Both of those things are true at once.

The stance that follows is neither to reject them nor to treat them as facts. Use them for the daily steering they are good for, know roughly how much of any figure they represent, reconcile in aggregate on a routine, keep a blended view that contains none of them, and settle the questions that decide real money with a test that measures rather than estimates.

Modelled figures are not confined to one number, which is part of why they are easy to miss.

Conversion counts are the obvious place, and usually the only one an account notices.

Conversion values are modelled too, so return on ad spend inherits the estimate.

Attribution across touchpoints involves modelling wherever journeys are incomplete, which means a data-driven figure is distributing partly estimated conversions across partly observed paths.

Audience sizes and reach figures are estimates by construction, since deduplicating people across devices is inference rather than counting.

Lift results from platform-run tests are computed on the same partly modelled data.

The cumulative effect is that a modern platform report is a blend of observation and estimate throughout, rather than an observed core with an estimated supplement bolted on. Knowing that changes how much weight any single figure should carry.

Three habits cover most of it.

Prefer figures that cannot be modelled when the decision is large. Total revenue, total spend, the customer count in the business's own systems, and the result of a test the business designed.

Treat platform reporting as directional at the level of individual campaigns and as approximate at the level of totals.

Improve observability rather than arguing about the estimates, because the only durable response is to leave less to be estimated.