Vision-language models describe images fluently, then fail at counting objects or judging spatial relations a child gets right. Is grounding a data problem, an architecture problem, or are we overestimating what the benchmarks prove?