About modelled conversions

Holdback validation (a machine learning best practice) maintains the accuracy of Google's models. A portion of observed conversions (validation data) are held back and split. Then, validation data that was run through the model is compared with validation data that wasn't. The validation results are used to check for inaccuracy and to further tune the model.

Modelled conversions are only included when there's high confidence of quality. If there isn't enough traffic to inform the model, then modelled conversions will not be attributed to ad interactions or, in the case of Google Analytics, are attributed to the 'Direct' channel. This approach allows Google to recover loss of observability while also preventing over-prediction.

Google's more general modelling algorithm is separately applied to your data to reflect your unique business and customer behaviour.

Google doesn't allow fingerprints or other attempts to identify individual users. Instead, Google aggregates data such as historical conversion rates, device type, time of day, geo and more to predict the likelihood of conversions from a specific ad interaction.

Some browsers, like Safari and Firefox, don't allow conversion measurement using third-party cookies. If you rely on third-party cookies for conversion measurement, you'll experience conversion modelling in line with your websites’ traffic on those browsers (desktop and mobile).

Some browsers, like Safari, limit the amount of time first-party cookies are allowed. You'll experience conversion modelling in line with your share of latent conversions beyond that window.

Regulations in some countries require that advertisers obtain consent for use of cookies related to advertising activities. Advertisers who have adopted consent mode will experience conversion modelling in line with their unconsented users. Conversions are modelled for unconsented users.

Apple's App Tracking Transparency (ATT) policy will require developers to ask for permission when they use certain information from other companies' apps and websites for advertising purposes. Google won't use information (such as IDFA) that falls under the ATT policy. In line with this, conversions whose ads originate on ATT impacted traffic will experience modelling. Make sure that your website can accept arbitrary URL parameters for the best modelling.

With the rollout of Apple’s ATT policy, SKAdNetwork, Apple’s app attribution solution has become an important input to app advertisers in assessing their iOS campaign performance. To improve the quality and consistency of Google’s modelled reporting in the Google Ads UI, we’re deepening our integrations with SKAdNetwork.

Google Play announced some new policy updates to bolster user control, privacy and security. As part of the Google Play services update in late 2021, the advertising ID will be removed when a user opts out of personalisation using advertising ID in Android settings. Any attempts to access the identifier will receive a string of zeros instead of the identifier. Learn more about advertising ID.

As a result of this service update, we'll be expanding modelled conversions to all App campaigns. This means that your conversion column as well as your install, in-app action and conversion value columns may contain modelled conversions. In the future, there may be additional modelled conversions in app campaigns as a way to mitigate impact that may result from this and other potential service updates.

When a user begins their journey on one device with an ad interaction and completes the conversion on another, it may not be possible to attribute the conversion to the ad interaction. Google observes data from the large number of signed-in users on Google properties to extrapolate similar behaviour across all users. Many cross-device conversions are also modelled, including from Living Room and Desktop.

Like all other products, Google's data scientists continuously make algorithm improvements to increase accuracy and scale of modelling. New products are regularly introduced to get new sources of observable data, which fine-tune Google's modelling. For example, enhanced conversions and consent mode can get more observed data.

Google uses techniques like holdback validation to check the accuracy of our modelling. For example, Google holds back a portion of observed conversions and models for that slice, then the modelled results and the actual observed conversions that were held back are compared, inaccuracies and biases are measured, and the models are continuously tuned. Similar methods are broadly used in Google AI.

Google only includes modelled conversions in their reporting when they're highly confident that conversions actually occurred as a result of ad interactions. Google avoids systematically reporting more conversions than reality and always aims to minimise over-reporting. This means that for some users, they don't observe enough conversions on a regular basis to be able to model accurately. In these cases, Google doesn't report any modelled conversions.

Because Google identifies different gaps in measurement and different types of observable data are needed and available, they have different types of models for different types of gaps. Google also uses techniques that eliminate double-counting across various types of models. Google knows that conversion rates vary significantly by advertising channel, and as a result, they build separate models for each channel and ad interaction type, that are impressions and clicks.

After a general modelling algorithm is determined to address a specific observation gap, Google applies that algorithm to each advertiser’s data separately and arrive at unique results that reflect unique user behaviour and conversion rates for that advertiser. For example, if your users have a very high tendency to start their journey on one device and convert on another device, there will be a higher than average cross-device modelled conversions reported for you.

For certain segments of traffic, Google will rely on additional signals to measure where conversions have occurred. These signals include, for example, the use of IP address to estimate conversions.

Google constantly runs experiments before rolling out any modelling changes, and if they detect a significant reporting and bidding impact, they communicate accordingly.

Where Google can accurately do so, they'll use available data to provide integrated conversion modelling in your conversion reporting and optimisation. In some cases, such as when conversions cannot be observed for a set of users that haven't consented to cookies, they'll need data about your consent rates so that they can provide conversion modelling.