This paper examines multimodal entity linking, which connects text and image mentions to knowledge-base entries, and finds that system performance degrades significantly on rare entities. The authors argue that prior rarity measures based on popularity metrics like pageviews are insufficient, proposing instead knowledge-graph structural metrics that better capture how well an entity is documented and connected. These structural metrics reveal many rare entities overlooked by popularity-based approaches, across which state-of-the-art accuracy drops substantially.