A brief dive into Retargetly's experience testing Google's Privacy Sandbox

Google announced the end of support for third-party cookies for its Chrome browser in 2024. To replace some of the core functionalities that third party cookies provide to the web, the Chrome team set forward the Privacy Sandbox and engaged the W3C and other industry players to get crucial feedback.

‍A mix of browser specific APIs make up Privacy Sandbox, and we decided to take a look at one of those for now: the Topics API. A browser-first opt-in API that categorizes users based on their past browsing behavior, the Topics API is the successor of the initial FLoCs API and one of the core advertising APIs inside Privacy Sandbox.

In September of 2022 we announced that we’ve partnered with the Chrome team as an early tester of the advertising APIs inside Privacy Sandbox. We’re doing this because we believe that as an industry we have to lean in and engage with the solutions that will, regardless of how we may feel about them, end up shaping user experience in the world’s number 1 browser. Our first analysis is done, and we’re ready to share some of the high-level findings.

We’ve approached our analysis from two perspectives. First, a lift and shift approach to answer how would using the Topics provided by the Topics API for behavioral targeting replace our current 8,000 segment taxonomy and what if any effect our clients would see.

Second, a real-life enrichment approach that uses the Topics to enhance existing anonymous customer profiles that are otherwise limited to a particular domain. For this approach, we ran our enrichment process over 48 days starting from January 1st 2023. In that time, we recorded Topics into the profiles of 10 million unique users with an average of 600,000 users per day.

A high level summary of our findings below:‍

‍• Audience Taxonomy: Topics v1 contains 349 IAB audiences, while Retargetly hosts 455 audiences in the affinity and in-market segments, similar to Google's offerings. Retargetly's full taxonomy includes over 5,000 audiences.

‍• Website Classification: Topics classifies websites based on their hostnames, resulting in a limited range of associated Topics and therefore, lower audience volume and category diversity. An improved classifier from Chrome is anticipated.

‍• User Profile Reach: Topics API yielded limited user profile data, with 55% of users generating no topics in the API calls over a 45-day study. Among users with some Topics, the average per call was 1.9, falling short of the theoretical 3 topics per call.

‍• Audience Comparison: Audiences based on Topics compared to those based on third-party cookies were frequently tenfold lower. Topics are currently accessible to 6% of Chrome users, but upgrades are underway to increase this to 100%.

‍• Response Time: The IFrame-based Topics retrieval method displayed a slow response time, 389 ms in the 90th percentile, hindering real-time bidding. A header-based approach is being assessed for potential improvements.

• Topic Rotation: For over 60% of users, half or more of their Topics remained constant from one week to the next. Google's designed topic rotation mechanism, based on weekly epochs, upholds user privacy but presents stability challenges for model creation.

‍• Machine Learning Application: Incorporating Topics into segmentation and identification models shows promise for a cookieless landscape. Tests revealed a 10% increase in reach when using first-party audiences with Topics as a signal.

Given our initial findings, our future research efforts around the Topics API will be centered in using it as an additional signal for our segmentation and optimization models rather than treating individual Topics as a segment.

2024 will definitely come with several changes in the industry, and the solution aims to combine a variety of tools in order to maximize our results. The use of Google Topics alone may not fully satisfy the requirements for interest-based advertising, but the combination of contextual features, first party data, Topics, FLEDGE, and attribution APIs could do the trick.

The Retargetly team that worked on this project: Pedro Zenone, Alejandro Rodriguez, Gonzalo Plaza, Nicolas Bisso, Matias Olejnik and Gustavo Pereyra.

###

Spanish: Una breve inmersión en la experiencia de Retargetly probando Privacy Sandbox de Google

Portuguese: Uma breve imersão pela experiência da Retargetly testando o Privacy Sandbox do Google