Modeled conversions use data that does not identify individual users to estimate conversions that Google is unable to observe directly. This can offer a more complete report of your conversions.
We model to recover slices of data where we know we cannot observe ad attribution due to protecting user privacy or technical limitations. We do this to provide high quality measurement so you accurately understand the impact of your marketing and maintain high quality bidding to prevent underbidding or overbidding.
When Google surfaces modeled conversions in Google Ads, we are predicting attributed conversions. In most cases, Google will receive ad interactions and online conversions but is missing the linkage between the two. The modeling we perform is modeling whether a Google ad interaction led to the online conversion, not whether a conversion happened or not.
Without modeling, reported conversions would only reflect the observable portion of conversions rather than the true campaign performance.
In order to model for a non-observed slice of data, we try our best to use data from observable slices where we know behavior is the same or very similar to the unobserved slice, or we have a good understanding of how they are different.
Example: Let’s say you have a slice of conversions that are not observable on one browser type, but can be observed on other browser types. Our modeling will first understand the trends between users' behavior (for example, conversion rates) across browser types. We then use our observable data from measurable browsers, together with any systematic biases, and incorporate other aggregate dimensions like device type, time of day, geographic location, operating system, and more, to predict the likelihood of conversion events from ad interactions on the unobservable browser type.
Modeled conversions are reported with the same granularity as observed conversions. This includes dimensions such as conversion totals, attribution path, and conversion values. In the “Conversions” column, Google reports both modeled and observed conversions.
Benefits of modeled online conversions
- Holistic measurement across all your ads traffic: Gain a more accurate picture of your advertising outcomes (ROI), and a complete picture of the conversion path across devices and channels resulting from ad interactions.
- Efficient campaign optimization: Modeled conversions help you optimize your campaigns more effectively and achieve better business results.
- Privacy regulations and technology limitations mean that we lose observation for certain cohorts of users (for example, unconsented users, or users using particular device types or browsers). This means our automated bidding algorithms will need to make optimization decisions based on incomplete data, resulting in biased learning. As a result, automated bidding may deprioritize those cohorts since they have a lower reported performance, leading to overall poorer performance by the bidder. Modeling solves for these biases and corrects them in overall reporting to ensure automated bidding has access to a more representative performance data. Learn more About automated bidding
- Accurate privacy centric measurement: Modeled conversions use data that does not identify individual users to estimate conversions that Google is unable to observe directly. This can offer a more complete report of your conversions.This approach is in direct contrast with non privacy-safe tactics like fingerprinting, which relies on heuristics, such as IP address, and attempts to identify and track individual users. Google has a strict policy against utilizing fingerprinting for ads personalization, as it doesn't allow reasonable user control and transparency.
Google's conversion modeling approach
Google solutions work across a broad array of users allowing the accuracy of our conversion models to be validated across a large set of ad interactions and conversion actions through several key dimensions:
- Scale: We have access to a magnitude and diversity of ad interactions across channels across various parts of the funnel. This provides us with comprehensive data around how different users react to different types of ads, regardless of where they are in the funnel and across all channels.
- Accuracy: Our high signed-in user base allows our sophisticated modeling techniques to operate independent of cookies or other identifiers since we can infer a rich set of behavioral dataset across a representative set of opted-in users.
- Coverage: Many websites use Google tags, which means that our conversion models are validated across a large set of different conversion actions. Conversion modeling uses data that does not identify the user in order to quantify conversions that Google is unable to observe directly. Our model is then trained uniquely on each advertiser, generating unique results.
- Technical Expertise: Google’s expertise in machine learning is a key capability allowing us to model with highest quality. We have mastered this throughout our measurement products that have employed modeling for years (Google Ads automated bidding and store visits) as well as products beyond measurement (for example, driverless cars and YouTube recommendations).
- Actionability: Google’s modeled conversions are surfaced in campaign reporting but are also tied to optimization and bidding. This renders the data actionable as it works towards your business goals.
Examples of available modeling for online conversions
Some of the most important conversion modeling efforts we have available are:
Modeling for third-party cookie limitations
Modeling for first-party cookie limitations
Modeling for EU cookie consent limitations
Impact of iOS 14
When a user starts 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 behavior across all users. All cross-device conversions are modeled.
Principles of online conversion modeling
Constant quality improvement
Like all other products, our data scientists continuously make algorithm improvements to increase accuracy and scale of modeling. We regularly introduce new products to give us new sources of observable data, which fine tune our modeling (for example, enhanced conversions and consent mode can give us more observed data).
Sophisticated techniques on checking for accuracy
We use techniques like holdback validation to check the accuracy of our modeling (for example, we hold back a portion of observed conversions and model for that slice). Then we compare the modeled results to the actual observed conversions that we held back, and measure inaccuracies and biases and continuously tune our models. Similar methods are broadly used in machine learning.
Rigorous thresholds for reporting
We only include modeled conversions in our reporting when we have high confidence that conversions actually occurred as a result of ad interactions. We avoid systematically reporting more conversions than reality and always aim to minimize over-reporting. This means for some users, we don’t observe enough conversions on a regular basis to be able to confidently model. In these cases, we do not report any modeled conversions.
Each gap is addressed via a unique modeling methodology
The outcome of each model is unique to your business and user behavior
Once a general modeling algorithm is determined to address a specific observation gap, we apply that algorithm to each advertiser’s data separately and arrive at unique results that reflect unique user behavior 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 modeled conversions reported for you.
Strict policy against fingerprinting
Communicating significant modeling changes
Where we can accurately do so, Google will use available data to provide integrated conversion modeling in your conversion reporting and optimization. In some cases, such as when conversions cannot be observed for a set of users that have not consented to cookies, we will need data about your consent rates so that we can provide conversion modeling.