Growth Notes

Modelled Conversions

Tags: analytics commerce concept

Modelled Conversions

Date: 2026-08-16

Vendors estimating the conversions that consent denial and blocking removed, then reporting them alongside observed ones. The estimate may be reasonable; what you can’t do is audit it, reconcile it, or separate it from what actually happened.

What it is

Modelled conversions are conversions a platform infers rather than observes, using patterns from the population it can measure to estimate the population it can’t.

observed conversions      420    measured directly
modelled conversions      140    estimated
                          ───
reported                  560    presented as one number

The mechanism, broadly: take the consented users where both the click and the conversion are visible, learn the relationship, apply it to the unconsented traffic where only the click is visible. [CHECK: each vendor’s method differs and none publish enough to reproduce it — treat any description as approximate.]

Why it exists

Not vendor mischief. A real problem:

Without modelling, reported conversions fall by whatever the loss rate is, and every channel looks worse in proportion to how measurable its audience happens to be. That’s a real distortion, and modelling is an attempt at correcting it.

What’s wrong with it

You can’t audit it. The method isn’t published in reproducible detail, so you cannot check whether the estimate is reasonable for your business, your category, or your traffic mix.

You can’t reconcile it. Modelled conversions don’t exist in your order system. So the gap between platform-reported and actual orders now contains both tracking loss and modelling, and you can’t separate them — Tool Discrepancies.

The estimator has an interest. The party estimating how many conversions its own advertising produced is not disinterested. That doesn’t mean the number is wrong; it means the incentive runs one way and you have no way to check.

It assumes the unmeasured behave like the measured. The core assumption, and the one most likely to fail. Consenting users are systematically different from non-consenting ones — that’s the same selection problem as Ad Blockers and Tracking Loss, and modelling from one group to the other inherits it.

In plain terms: a share of the conversions in your ad platform never happened as recorded events anywhere. They’re an estimate produced by the company selling you the ads, using a method you can’t inspect.

Working with it

  • Find the split. Most platforms disclose modelled versus observed somewhere, even if the default view blends them. Get the split before treating any figure as a count
  • Never reconcile blended figures against your order system. Compare observed-only against your own tracking; treat modelled as commentary
  • Watch the modelled share over time. A rising proportion means measurability is degrading, which is a real signal about your tracking even if the estimate isn’t
  • Use it for within-channel decisions — creative A versus creative B, where the same modelling applies to both sides and largely cancels
  • Never use it for between-channel budget decisions. Different platforms model differently, so cross-platform comparison compares two different estimation methods — Walled Garden Reporting
  • Size channels with a holdout where the money is material — Incrementality Testing, Geo Holdout Tests

The position that holds

Modelled conversions are a vendor’s estimate of their own effectiveness, presented as measurement. Treat them the way you’d treat any supplier-reported performance figure: useful for direction, not admissible as truth, and never the basis for moving budget between suppliers.

The alternative isn’t better modelling. It’s a holdout, which answers the question the modelling is trying to approximate — and answers it with your own data.

Graph View

Backlinks