Lost in Embeddings: Information Loss in Vision-Language Models