Selecting an effective cover photo is a critical yet understudied problem in visual merchandising for e-commerce marketplaces. On Airbnb, hosts upload many photos depicting different rooms, amenities, and points of interest, but must manually choose a single static cover photo that is reused across heterogeneous product surfaces. This creates substantial challenges: over 90% of listing photos have no engagement labels, leading to a severe cold-start regime; guest preferences vary across surfaces and search intents, and visually captivating photos may be semantically misaligned with the underlying listing. At Airbnb scale, this corresponds to hundreds of millions of photos and hundreds of thousands of new uploads per day. We propose OptiCover, a scalable framework for cold-start visual optimization under semantic constraints and multi-surface engagement signals. The framework combines (i) a photo candidate generator that identifies candidate photos through various proxies, (ii) lightweight vision heads applied to frozen CLIP embeddings for large-scale semantic scoring, (iii) a cross-modal grounding validator based on a vision-language model (VLM) that enforces semantic representativeness, and (iv) an Exploration Logging Module (ELM) that orchestrates the collection of randomized engagement data across heterogeneous product surfaces and contextual search intents. We evaluate OptiCover through extensive offline studies and multiple online experiments across HomePage, and Search surfaces. OptiCover is now deployed in production and performs tens of millions of photo swaps daily. Online results show a 1.5% engagement lift on the HomePage and a 0.35% increase in bookers on Search, generating millions of dollars in incremental revenue. Our findings suggest that action-space shaping and semantic constraints are primary drivers of performance.
Figures & tables
Figure 1. Airbnb Search with Pool as filter before and after applying OptiCover Airbnb Search with Pool as filter
Figure 3. Vision heads framework Vision heads framework Diagram A diagram showing the Vision heads framework with CLIP embeddings feeding into the quantizer, then undergoing either model training (classification or pairwise classification) before uploading the weights into the model store
Task
Metric
Baseline
Head (Ours)
Room type*
macro F1
0.54
0.96
Annotation
macro F1
0.53
0.84
Human in photo
macro F1
0.66
0.92
Amenities
macro F1
0.65
0.90
HomePage CTR
Pairwise acc.
57% (Host)
68%
Table 1. Vision heads performance, baseline is zero-shot unless specified otherwise, * indicates a multi-label task
Source
Description
Current Host Cover
The currently selected host cover
Labeler Selected Cover
A photo selected by an internal labeler as part of the Airbnb Categories launch ( airbnb22 )
Past Host Cover
A cover photo used in the past
Similar
A photo semantically similar to an already vetted candidate
Opportunistic Cover
A photo with high engagement improvement ( ΔCTR^×weekly_traffic )
Table 2. Sources of photo candidates
Data
Details
Photo
The image content of the listing photo.
Caption
Host-provided caption, which may include contextual or descriptive voice-over (e.g., “This tavern is a 5-minute walk away”).
Listing description
Host-provided description of the space, including features, amenities, and nearby points of interest.
Table 3. Listing grounding validator input
Surface
Duration
Aprx. N
CTR lift
Bookers lift
HomePage
3 wks
∼ 10M
+1.5% [1.3%, 1.8%]
–
Search (Contextual Exploit)
6 wks
∼ 1M
+0.14% [-0.24%, 0.52%]
+0.35% [0.01%, 0.69%]
Search (non-contextual)
2 wks
∼ 10M
+0.32% [0.16%, 0.48%]
+0.075% [-0.14%, 0.29%]
Table 4. Online experiment results. N denotes the approximate number of unique visitors. Values report relative treatment effects with 95% confidence intervals. Bold indicates statistical significance at α=0.05 .
Candidate photo set added (cum.)
Best CTR (normalized)
Photo count (normalized)
Current Host Covers
1.00
1.00
+ Labeler Selected Covers
1.10
1.19
+ Past Host Covers
1.23
1.34
+ Similar
1.34
1.47
+ Opportunistic Covers
1.38
1.51
Table 5. CTR of best performing photo improves with additional candidates
Figure 4. Comparison of in-training auxiliary calibration loss (left) and post-training calibration (right) on a held-out set. Calibration comparison
Figure 5. Regret gap between Thompson Sampling and CTR Exploit over the online experiment. Positive values indicate better performance for CTR Exploit.
Harbin Institute of Technology, Shenzhen, China · Peng Cheng Laboratory, Shenzhen, China · Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China.