Decompiled Android Packages Expose 527-Sound Tracking

Google's user tracking extends beyond NavBoost SERP click logs into local models that classify physical activity and ambient audio on the phone.

Olivier de Segonzac reports decompiling Android packages such as Personal Intelligence and AICore, which expose endpoints that download and run behavioral models on the device.

Unlike standard web analytics on search results, these local models detect whether a user is walking, running, cycling, riding as a vehicle passenger, taking a bus, metro, or train, skiing, flying, or boating.

Because the microphone listens continuously for the "OK Google" wake phrase, the phone also runs an embedded AudioSet ontology of 527 sound types with threshold-based triggers — recognizing an alarm, a crying baby, laughter, a dog barking near home, whistling, a slamming door, or applause, which triggers no action.

Google builds a map of recurring weekly habits, such as a Thursday evening workout after work — loading the user's usual music playlist — or a Saturday morning grocery run.

Those detected routines feed Google Discover cards and generate Google Search zero-prefix suggestions, populating personalized query prompts the moment a user places the cursor in the search bar before typing a single character.

Separately, Google embeds local intent classifiers inside Chrome — inspectable via chrome://components and Chrome Canary — to keep processing on the device and cut server load.

While full local models such as Nano 4 reach 4GB across 1 billion to 2 billion Chrome users, Google distributes lighter opt-in models including Embedding Gemma 2 at ~250MB alongside a TypeSafe- and JEEVES-style classifier deployed on October 8 — just 15 days to 3 weeks after JEEVES launched on September 15.

Running on the user's CPU via LoRA fine-tuning weight files, the classifier scores page structure against a predefined JSON category taxonomy and adapts browser features by page type, often rolling out in beta on US sites first.

On content sites, a top-of-page Gemini bar injects pre-generated prompts offering to summarize the editorial article, whereas on shopping and e-commerce sites it prompts the user to synthesize customer reviews.

On travel and booking platforms such as Booking.com, TripAdvisor, and Hotels.com, Google is currently working on getting local models to parse interactive form schemas and translate natural-language requests — such as a stay in 3 weeks for 3 people and a dog at a beachfront hotel — into pre-filled form fields, letting the user launch the search without manual on-page navigation.