Google’s Privacy Sandbox APIs Explained: Topics API

The Privacy Sandbox is a new set of features in Chrome that allows advertisers to reach potential customers and publishers to monetize their content while preserving user privacy. One of the new features is the Topics API which enables advertisers to target users based on their interests. This article will briefly explain how the Topics API works and how we are experimenting with it at NextRoll.

The Topics API observes a user’s browser history and records the topics the user is interested in based on the domains in their browsing history. Each domain is associated with one of 470 unique topics, such as Arts & Entertainment or Shopping/Children’s Clothing. The domain classification scheme is hand-curated by analyzing the top 50,000 domains and derived by machine learning models for domains outside this list. The Topics API enables a supply side partner (SSP) to ask the browser for a random selection of the topics a user has most frequently visited in the past few weeks.  As a note, this selection can only include topics from publishers that work with that particular SSP, so the Topics API favors larger SSPs that work with more publishers. 

At NextRoll, we are currently running two experiments testing this functionality to better understand how we can leverage user interests to optimize ad performance. The first experiment is an A/B test where our bid value for our auctions test group depends on the topics those users are interested in. In the bid request, we receive topics that a particular SSP has observed for each user. We then bid in proportion to the likelihood each of these users will click on an ad related to those interests. These results will help us learn if our advertisers who spend more budget on topics with high click probabilities will see better performance than those who spend on low click probability topics. While it’s too early to share our performance analysis, initial results indicate there is little variation in click probabilities across different topics. This is an intuitive result: the likelihood that a user clicks on an ad should not solely depend on what topics they’re interested in, but rather a mix of their interested topic AND the content of the ad being shown. 

In our second A/B test, we are comparing performance of our standard bidding algorithm to a new algorithm that uses topics, but excludes granular user features. For the control group, our standard algorithm selects bid prices based on a variety of signals, such as the domain of the bid request, advertiser features, and granular user features collected via third-party cookies. For the test group, we’re using a new algorithm that uses all of the same signals as our standard algorithm, but excludes the third party cookie data. Instead, we’re using topics that the SSP passes to us through the bid request. Since these algorithms require a lot of training data, this test is running through off-device auctions instead of on-device auctions. Although on-device auctions have yet to reach the scale required to produce meaningful results, we hope to learn if topics are a suitable replacement for third-party cookie data. 

We are looking forward to analyzing the results of our experiments and letting those results guide our next steps with the Topics API. NextRoll believes that privacy-forward advertising will be an important part of the future web, and we are actively testing all features of the Privacy Sandbox. 

Alex Verbny is a Staff Data Science Engineer at NextRoll