Privacy-safe conversion modeling fills in the unknowns in the customer journey.
Conversion modeling uses observed conversions to predict unobserved conversions without identifying any one individual.
| Observed conversions | Modeled conversions |
|---|---|
|
Uses cookies and other identifiers to link between ad interactions and conversions. |
Uses machine learning to assign links between ad interactions and conversions accounting for cases where cookies and identifiers weren’t available. |
When are modeled conversions included in your reporting
Modeled conversions are included in your total reported conversions only when there’s a high confidence that your ad resulted in conversions. This rigor ensures that we avoid systematically over-reporting. And in cases where we don’t have enough data to be able to confidently model, we don’t provide conversion modeling.
Holdback validation (a machine learning best practice) maintains the accuracy of Google’s models. Modeling methodology is applied to a subset of observed conversions, to understand the accuracy of the model by comparing to observed results. This info is used to fine tune the models.
Google constantly runs experiments before rolling out any modeling changes, and if we detect a significant impact on your data, we communicate accordingly.
Conversion modeling works like this:
1. Ad interactions are separated into two groups
One group contains ad interactions that have a clear, observable link to a conversion. The other group contains ad interactions that don't have a clear, observable link to a conversion.
2. The observed group is divided into subgroups
The observed conversions are divided into subgroups based on shared characteristics, and key metrics are calculated for each. For example, conversions observed in the morning in France are found to have a certain conversion rate, whereas this rate may be different in the evening.
3. The unobserved group is sorted into those same subgroups
Those subgroups are used for sorting unobserved ad interactions and conversions.
4. Unobserved ad interactions and conversions are linked
Using the known conversion rates and other characteristics from the observed subgroups, machine learning links unobserved ad interactions and conversions, where appropriate. The observed and modeled conversions are then integrated into your conversion data to help you make informed decisions about ad performance reporting, and fed into bidding to provide an unbiased view of your performance. This leads to better optimization.
In practice, calculations from observed data are based on a variety of dimensions, including location, time, and browser. These are combined with data from platform APIs, surveys and user panels to further refine modeling.
Privacy-centric approach
Google doesn’t allow fingerprinting or other attempts to identify individual users. Instead, Google uses aggregate data (such as historical conversion rates, device type, time of day, geo, and more) to predict the likelihood of conversions by users who viewed or clicked on an ad.
