Topics API

The browser performs all topic classification locally on the user's device, without transmitting raw browsing history to any external server. The browser maps visited hostnames to a fixed, human-curated taxonomy of approximately 350 topics, such as 'Fitness' or 'Travel'. This ensures that granular behavioral data never leaves the device, fundamentally shifting the privacy model from server-side profiling to client-side computation.

Topics are calculated on a weekly epoch cadence. For each epoch, the browser selects one topic per week from each of the past three weeks to share with callers. The topic for the current epoch is derived from the user's browsing activity during that week. This temporal batching prevents real-time tracking and limits the granularity of the interest signal, making it impossible to infer a user's minute-by-minute browsing behavior.

The API relies on a static, public taxonomy of human-readable topics rather than arbitrary, opaque segments. Key characteristics:

• The taxonomy is maintained in a public GitHub repository
• Topics are intentionally broad and non-sensitive (e.g., 'Automotive' not 'Luxury Sedan')
• Sensitive categories like race, religion, or sexual orientation are explicitly excluded
• Site-to-topic mapping is performed using a manually curated override list and a lightweight ML model

Each calling context (the embedding site) receives only the topics it has previously observed from the user. If a caller has never seen a particular topic for a user, that topic is not returned. This prevents cross-site data leakage: an advertiser cannot learn about topics inferred from a publisher they have no relationship with. The browser maintains separate observation sets per caller origin.

The API introduces controlled randomness to prevent re-identification. With a 5% probability per epoch, the browser returns a completely random topic from the full taxonomy instead of the true inferred topic. This differential privacy mechanism ensures that even if an observer collects topics over many weeks, they cannot be absolutely certain that any single topic reflects genuine user interest.

Users have direct visibility and control over their topics. The browser provides a settings interface where users can:

• View the topics currently associated with their browsing
• Remove individual topics they do not wish to share
• Opt out of the Topics API entirely This contrasts sharply with opaque third-party cookie tracking, where users had no practical insight into the segments assigned to them.