Running RTGT campaigns on iOS is challenging due to the reduced number of users sharing their device level information. When we receive bid opportunities, a significant portion of them lack a device ID, making it difficult to identify users. This presents a specific challenge for RTGT, as we aim to target a defined user segment rather than the open market. How can we achieve this without relying on device IDs?

For some advertisers, we have tested an alternative method of identifying users that does not rely on device IDs but instead leverages IP address information available during the auction. For clients like Rakuten and Bigo, we created specific RTGT audiences based on the same logic and criteria as regular audiences, but with IP addresses as the output.

Using IP addresses alone presents several limitations, mainly because it is not always accurate to assume that the user behind an IP is the intended target. Frequently, multiple users share the same IP address: for example, in a family household, each member may use different mobile devices but share the same wifi connection and therefore are grouped under the same IP. This problem is amplified in public spaces like restaurants or shopping malls where many users connect through a common IP address. Therefore, this approach is not scalable and would not consistently deliver the performance we aim for.

To address these limitations, we developed a more sophisticated framework: the "Jampp Probabilistic Identifier". This identifier is constructed by us using available data—excluding the device ID—to probabilistically infer the user we want to target. Our first iteration of the "jampp_probabilistic_id" incorporates IP information, OS, and OS version.

By adding OS and OS version data, we improve the precision of our identification. These data points are relatively easy to access and help us refine our understanding of the user behind an IP address. However, the true potential of the Jampp Probabilistic Identifier lies in its adaptability: we can continue to integrate additional data points to improve accuracy and reduce the "confusion rate"—the error margin between the initial data point (e.g., IP) and the actual user we want to target (more on this on section 8).

To validate that a non-IDFA Probabilistic Audience can be set for the groups within your campaign, you have to carry the following checks before:

Notes on each point:

In Silver, there are two main new differences:

In the audience form, you’ll find a checkbox labeled “Probabilistic”. Selecting this option works similarly to creating a regular audience, with the only difference being that it will also bid for non-IDFA traffic—that is, traffic without a device ID.

What does this mean? There’s no need to create two separate audiences—one for device ID traffic and one probabilistic—if the targeting conditions are the same. You simply need to activate the “Probabilistic” checkbox to allow the audience to run for non-IDFA traffic as well. However, be sure to read the next section on the “Group Form” to understand how to configure it to bid only for non-IDFA traffic if needed.

In the group form, under the “Audiences” section, you’ll find a new “Bidding traffic type” menu, which determines whether the group bids for both IDFA and non-IDFA opportunities or just one of the two. The available options are:

Best practice: it’s recommended to create two groups—one for IDFA-only and one for non-IDFA-only. This approach allows you to control how much budget is allocated to each strategy, and ensures that each group follows its own specific pacing needs.

The primary objective of using IP addresses initially—and now the jampp_probabilistic_id—is not to create an audience that outperforms those we can already build. Instead, the goal is to unlock a portion of traffic that was previously inaccessible due to the absence of device IDs. Until now, when using an RTGT audience, if an opportunity came without a device ID due to iOS opt-in rates, we simply would not bid for that user, resulting in the loss of a significant number of potential conversions.

Although the jampp_probabilistic_id is not perfectly accurate—and improving its precision is a core part of our roadmap—it is better to have this mechanism in place than not: while the current accuracy can be improved, this feature already allows us to access a previously unreachable audience segment.

While OS and OS version improve the accuracy, we recognize that the most valuable differentiating factor is the device model. Users sharing an IP address are likely to have the same OS version (e.g., iOS 17.5.1), but it is less probable that they all use the same device model (e.g., iPhone 14, iPhone 15, iPhone 15 Pro). Among different datapoints, we see device model as the key piece of information for refining the Jampp Probabilistic Identifier.

However, incorporating device model requires additional time and effort: we need to align how this field is reported across various data tables—primarily auctions and events—before it can be integrated into the identifier.

Since the jampp_probabilistic_id and probabilistic audiences are designed to allow us to target non-identified users that were previously inaccessible, the primary measure of success is not whether these audiences immediately reach better performance than deterministic regular audiences. Instead, our focus is on continuously improving the accuracy of the probabilistic audience to reduce the gap between the targeted user and the actual user we aim to reach.

As we transition from using IP-only audiences to the jampp_probabilistic_id, our objective is to reduce the "confusion rate"— as mentioned above, the margin of error between the initial data point (e.g., an IP address) and the actual user we intend to target. Initially, targeting by IP alone is not precise because multiple devices may share the same IP address (as illustrated by the shopping mall example). By incorporating additional information such as OS and OS version, we can begin to differentiate between devices sharing the same IP, therefore reducing the confusion rate. In a similar way, integrating the device model information will also further refine our ability to distinguish between devices: while a group of users under the same IP is likely to share the same OS version (e.g., iOS 17.5.1), they are less likely to all have the same device model (e.g., iPhone 14, iPhone 15, iPhone 15 Pro).

In addition, there are other alternative audience cases which might or might not support the use of non-IDFA Probabilistic Audiences: