In user-generated marketplaces, sellers are unlikely to provide structured text descriptions, making image embeddings the primary available signal for item similarity. Visual neural embeddings enable the recommender to identify related items and surface relevant alternatives even when textual metadata is sparse or absent. This directly supports recommendation coverage across the full catalog rather than only well-described listings. For a marketplace operator, this translates into broader monetizable surface area - more items receiving recommendation-driven impressions - without requiring sellers to change their listing behavior. The business case is strongest where catalog quality is variable and text-based signals are unreliable.