When a platform cannot observe all conversions, it estimates them with statistical models. These numbers appear in reports as if they had been measured. Understanding what they are, when they weigh most and how to validate them helps you read results without illusions.
The essentials in a definition
Conversion modelling is the platforms’ use of statistical methods to estimate conversions they cannot observe directly, for example when the user refuses cookies. It does not identify people; it uses aggregated patterns from those who accepted. It depends on correct implementation of consent mode and the developer guidance, and on minimum volume.
What modelled conversions are
They are conversions the platform infers from patterns in situations where there is no direct data: refused consent, browsers limiting identifiers, device switches. The model estimates how many conversions would have occurred and adds them to the report.
Unlike a sale recorded in your system, a modelled conversion is an aggregate probability, not an individually verifiable event.
When they weigh most
In markets with a lot of refused consent, on sites with many visitors using restrictive browsers, or in campaigns with many cross-device conversions. The modelled fraction may be small or large, depending on context.
Platforms do not always indicate the percentage. When they do, it is worth tracking it over time.
Reading modelled conversions
- Know what share of the total is estimated
- Compare with real sales by period
- Cross-check with tests and revenue
- Improve direct observation
How to validate
Compare the platform’s total conversions with real sales, by period. If the ratio stays stable, the model seems to track reality. If it drifts, investigate configuration, consent and possible duplication.
Use incrementality tests and revenue data to confirm that modelled conversions correspond to sales.
A modelled conversion is an honest estimate, not an invoice.
How to decide with this kind of data
Treat it as a useful estimate for day-to-day optimisation, not as accounting. For larger budget decisions, cross-check with other sources. Avoid reporting to the client or management as exact numbers.
Where possible, improve observed data: well-designed consent, enhanced conversions and server-side measurement reduce reliance on models.
Decision table
| Aspect | What happens | Implication |
|---|---|---|
| Those who consent | Data observed normally | Basis for modelling |
| Those who refuse | Limited signals, no identification | Estimated conversions |
| Volume | Minimums needed to model | Small accounts may have no modelling |
| Reports | Mix of observed and estimated | Understand which is which |
Common mistakes to avoid
- Thinking modelling is covert tracking. It does not identify individuals; it respects the choice to refuse.
- Misconfiguring consent. Without correct signals, modelling does not work and there is legal risk.
- Reading it as an exact number. These are estimates with uncertainty.
How to measure the result
Track the consent rate, the proportion of modelled conversions and the ratio between recorded conversions and real sales before and after enabling modelling.
Limits and exceptions
Conversion modelling is an estimate and is not available under all conditions: it requires minimum volume, correct implementation and retains uncertainty. It does not correct tracking failures or replace first-party measurement. Modelled values may be revised by platforms. Track the difference against real sales and avoid treating the estimated number as exact in important decisions.
Illustrative example
Illustrative example: a site with a 55% consent rate finds observed conversions are about half of sales. After correcting the consent implementation, modelled conversions begin to be included in reports and the gap against real sales narrows. The company keeps comparing with the sales system, and does not take the estimated value as exact.
Frequently asked questions
Are modelled conversions reliable?
They usually track reality well in aggregate, but should be validated with your own data.
Can I switch them off?
Generally not. They are part of the platforms’ measurement system.
How do I reduce dependence?
With better consent, enhanced conversions and server-side measurement.
Does modelling identify those who refused?
No. It uses aggregated patterns from users who consented to estimate the total, without identifying individuals.
Do small accounts get modelling?
It depends on the platform’s minimum volume; check the documentation for current requirements.
Conversion modelling estimates, in aggregate and without identifying people, what refused consent prevents observing; it needs correct implementation and should be read as an estimate.
Sources and recommended reading
- Analytics Help: consent mode
- Google for Developers: set up consent mode
- Google Ads Help: conversion measurement
Last editorial review: October 2026. Informational content; always confirm current conditions in the official sources before deciding.