Platforms like Zillow, Instagram, and Airbnb use images to compete for users’ attention and money. But what makes a photo stand out, and does a distinct image win more consumers?
After studying Airbnb listings in New York City, Harvard Business School’s Shunyuan Zhang and fellow researchers found that properties with less unique photos tend to have lower occupancy rates. However, being too original can scare consumers away.
The analysis, using nearly 482,000 images collected from April 2022 to April 2023, identified the optimal level of uniqueness, potentially translating into thousands of dollars more in annual revenue for the host. At scale, gains could also be significant for the platform.
The findings appear in “Visual Uniqueness in Peer-to-Peer Marketplaces: Machine Learning Model Development, Validation, and Application”, published in the December issue of the Journal of Consumer Research. Zhang, an associate professor, coauthored the research with Xiaohang (Flora) Feng, a doctoral student at Carnegie Mellon University, and Charis X. Li, an assistant professor at Grenoble Ecole de Management in France.
Despite its importance, assessing the visual uniqueness of an image, especially on peer-to-peer marketplaces, has been difficult to quantify at scale.
“The core challenge is that visual uniqueness is a subjective human judgment, and people can differ substantially in what they perceive as ‘unique.’ As a result, large-scale human evaluation can be costly, noisy, and inconsistent,” Feng explained. “Moreover, artificial intelligence approaches that rely on human-labeled notions of uniqueness may inherit these limitations and can be difficult to generalize across contexts.”
How can an algorithm assess uniqueness?
The authors drew inspiration from research on human visual perception, particularly how people process complex, unfamiliar images from the ground up, starting with visual features rather than preconceived categories.
The researchers developed a model that processes raw visual data at the pixel level. This approach, called unsupervised contrastive learning, does not rely on pre-labeled data. Instead, it captures patterns in color, texture, contrast, and the arrangement of elements. The model is trained by comparing image pairs: it pulls similar images together while pushing dissimilar ones apart. The paper compares this to sorting apples from oranges in a mixed fruit basket.
When applied to 61,000 Airbnb images that weren’t used for training, the model was able to accurately gauge visual uniqueness 73% of the time. It not only assigned a uniqueness score to each of the 481,747 images within the dataset, but also generated a heatmap showing how different pixel regions and visual features contribute to an image’s perceived distinctiveness.
The paper treats uniqueness as distinct from aesthetics. It doesn’t measure if a photo is good or bad, but studies how straying from category norms impacts consumers, Feng explained.
The researchers also tested how removing certain elements from an image—such as a zany rug or retro lamp—affects its uniqueness score. While removing objects can tone down an image, the approach can also backfire: eliminating certain components (blue in the heatmap) increases the uniqueness score, but can also raise red flags.
How platforms can leverage uniqueness
The findings by Zhang and her coauthors stand to benefit entrepreneurs and businesses operating in any context where images matter, including peer-to-peer marketplaces, online retail, and even social media. Based on their research, platforms might consider:
Identifying the details that influence customer behavior. Uniqueness can attract consumers, but it has its limits. Since uniqueness is not always better, platforms like Airbnb can use these insights to promote listings that are interesting yet not so unusual as to feel risky.
Adding tools that gauge uniqueness. A visual uniqueness measure could be incorporated to enhance search and recommendations, helping consumers discover listings that match their level of comfort, which in turn can create a better user experience.
Targeting listings at the extremes. Listings that are very off balance could benefit most from even minor changes, since the curve is steepest away from the optimum, the paper notes. Near the “sweet spot” for uniqueness, alterations don’t matter as much, but they do for listings that are too bland or excessively unique.
Hosts, particularly those operating multiple properties, could also gain from understanding uniqueness. Scores and heatmaps can help guide renovations, décor, and photo selection. Even simply removing the least unique image from a set increases the average uniqueness across listings, boosting effectiveness.
And when a zebra-print carpet or a lack of pillows puts off potential customers, host responsiveness can mitigate uncertainty.
“Hosts offering highly unique offerings should strive to provide quick responses and quality guarantees to mitigate perceived uncertainties for prospective guests and reap the maximum benefit of uniqueness,” the authors write.
Visual Uniqueness in Peer-to-Peer Marketplaces: Machine Learning Model Development, Validation, and Application
Feng, Flora, Charis Li, and Shunyuan Zhang. "Visual Uniqueness in Peer-to-Peer Marketplaces: Machine Learning Model Development, Validation, and Application." Journal of Consumer Research 52, no. 4 (December 2025): 800–825.
