A system matches Airbnb cover photos to guest intent
OptiCover: A Scalable and Multi-Surface Framework for Cold-Start Visual Optimization under Semantic Constraints · arXiv:2611.12727
In plain words
Airbnb hosts choose one cover photo, but guests use different surfaces and search for different things. More than 90% of listing photos have no direct engagement labels, and a photo that gets clicks may not represent the listing.
OptiCover finds and checks candidate photos, then uses guest responses to guide which eligible photo to show. In one Search test, matching photos to seven selected filters increased bookings; the measured click rate did not significantly change.
The research
A photo that earns clicks may still fail to help guests find a listing that fits what they are searching for. These results change the expectation that more clicks alone mean more bookings: in the reported Search tests, booking gains appeared when photo choices matched selected search filters.
OptiCover generates, checks and ranks Airbnb cover-photo candidates, then tests whether selecting by guest context improves engagement and bookings.
In randomized online tests against current host-selected covers, OptiCover increased HomePage click-through rate. Search results differed: applying a HomePage-trained click model directly raised clicks but not significantly Bookers, while a context-matched Search test increased Bookers.
Together, these results suggest that click gains alone may not translate into more bookings, and that matching eligible photos to search intent can matter. The Search tests also show that HomePage engagement data can start optimization on another surface.
The online results cover HomePage and Search experiments; the contextual Search test included seven selected filters. OptiCover cannot change listing-card order, so it only changes which photo appears within a card.
What the evidence shows
- Build choices and check fit
OptiCover is a framework for selecting Airbnb cover photos when most listing photos lack direct engagement labels. It generates and scores candidate photos, checks whether they represent the listing, and filters them for the guest-facing surface or search context.
An online component also gathers randomized guest responses to help learn which eligible photos draw engagement.
- More candidates, higher best CTR
An offline HomePage analysis asked whether expanding the available candidate set gave each listing a higher-CTR photo to choose. The normalized best-photo CTR was 1.00 with current host covers alone and 1.38 after all listed sources were added.
That supports the value of expanding the available choices in this analysis, but it is an offline best-photo comparison, not an online result.
- A check catches misleading photos
In the sampled candidates, recall of non-representative photos was 98% with the full pipeline and 89% without the grounding validator. This supports the validator’s role in catching misleading candidates in that sample; it does not by itself measure guest outcomes.
- HomePage clicks increased
In a three-week HomePage experiment, visitors were randomly assigned to OptiCover or the current host-selected covers. CTR, the share of impressions that result in clicks, was 1.5% higher with OptiCover relative to the control.
This online result shows an engagement increase on HomePage, where the system selected from eligible candidates using predicted CTR.
- Search intent changed the result
Search results depended on how candidate photos were selected. Applying the HomePage-trained CTR model directly raised Search clicks, but did not significantly increase Bookers, Airbnb’s measure of people who make bookings.
When a separate test constrained photos to match seven selected search filters, Bookers increased even though the CTR change was not significant. The comparison suggests that matching photo choices to search intent can matter for bookings beyond maximizing clicks.
Deep
- Original paper authors
- Guillaume Guy, Mihajlo Grbovic, Joe Mills, Selen Uguroglu, Liwei He
- Submitted to arXiv
- License
- arXiv non-exclusive