Airbnb Personalizes for Logged-Out Users Using Neighborhood-Level "Proximity Features"

Airbnb's engineering team describes how it personalizes search and homepage content for visitors it knows nothing about, a useful signal of how the platform treats cold-start traffic under tighter privacy expectations.

The system, called Proximity Features, serves users without login history, prior searches or persistent identifiers. Rather than tracking individuals, it groups geographically nearby users and aggregates their collective behavior.

How it works

• A "proximity key" represents a local cluster of about 1,000 users, combining quantized latitude/longitude tiles with IP hash buckets in dense areas.
• An adaptive two-phase clustering approach uses fine-grained tiles in dense urban areas and coarser tiles in sparse regions until each bucket reaches roughly 1,000 users. The partition bootstrapped in 2023 has stayed valid in production without re-clustering.
• Daily features cover short-term engagement (recent destinations, room types, median prices), long-term booking patterns (booked destinations, travel party signals) and aggregate metadata (bucket size, geographic density).

Results

Production A/B tests showed gains on marketing landing pages, which previously served static cards, and on homepage AutoSuggest, where location-aware suggestions replaced generic global defaults. Never-booked and dormant users showed directional gains, and destinations such as Kuala Lumpur, Dubai and Jeju Island appeared consistently in treatment groups.

Privacy design

All features reflect about 1,000 users collectively, clustering excludes non-consenting users, geo-IP coordinates stay coarse and group-level, and deletion controls plug into governance systems.