How Airbnb Rewired Search Ranking for Maps: Three Attention Models That Boosted Bookings

Introduction

Search is Airbnb's core mechanism for connecting guests with hosts. After a guest searches, results appear in two interfaces: (1) a list of rectangular cards showing photos, price, and ratings (list-results), and (2) elliptical markers on a map displaying prices (map-results). Initially, both interfaces used the same ranking logic: sort listings by predicted booking probability and show the top-ranked ones.

However, the assumptions behind list ranking do not hold on a map. In a list, user attention decays as they scroll down (Figure 1 shows CTR dropping sharply with rank). The algorithm therefore places higher-booking-probability listings higher to capture more attention. On a map, markers are scattered across an area; there is no ranked list and attention does not decay with rank position. Simply sorting by booking probability no longer works.

Uniform Attention Model

The first iteration assumed user attention is uniformly distributed across all visible map markers. Users click only a few markers (Figure 3). Showing too many markers means users may miss the best options within their limited clicks; showing only the very best risks excluding the listing a user would actually choose.

To test this, the team introduced a parameter that caps the ratio between the highest and lowest booking probability among displayed markers. Stricter caps raise the average booking probability of shown markers. Figure 4 summarizes A/B test results across a range of parameter values. Two key metrics improved as the cap tightened: "average impressions to discovery" (fewer markers viewed before the booked listing) and "average clicks to discovery" (fewer clicks needed). The version with the stricter filter delivered a booking lift ranking among the top improvements in Airbnb search history, and also increased high-quality bookings — more 5-star trips post-stay compared to control.

Tiered Attention Model

The next iteration split markers into two tiers. Highest-booking-probability listings appear as regular elliptical markers with prices; lower-probability listings appear as smaller "mini-pins" without prices (Figure 5). By design, mini-pins attract far less attention — their CTR is roughly one-eighth that of regular pins.

This approach is especially useful on desktop, where the left panel shows a fixed grid of 18 results, each requiring a corresponding map marker (Figure 6). Since the marker count is fixed, the team could not simply reduce the number of markers. Tiering instead directs attention toward the highest-probability markers. Figure 7 shows the A/B test results confirming the effectiveness of this two-tier design.

Decaying Attention Model

In the final iteration, the team mapped CTR at different map coordinates to understand how attention decays spatially. Figure 8 shows heatmaps for mobile (top) and desktop (bottom). Attention is highest near the map center and falls off toward the edges.

To exploit this, they designed a recentering algorithm that shifts the map center toward the highest-booking-probability listings. The algorithm evaluates a set of candidate center coordinates and selects the one that minimizes distance to the top-probability listings (Figure 9). In an online A/B test, this change lifted non-canceled bookings by 0.27% and reduced map moves by 1.5%, indicating users needed fewer interactions to find what they wanted.

Conclusion

User interaction with maps differs fundamentally from list interaction. By progressively refining models of user attention on maps — uniform, tiered, and spatially decaying — Airbnb improved real-world guest experience. A remaining challenge is how to surface all available listings on the map, which is part of ongoing work. Full technical details are published in the KDD 2024 paper "Learning to Rank for Maps at Airbnb".