We introduce a mathematical framework and benchmark to quantify generative-model steerability, reveal large steerability gaps in LLMs and text-to-image models, and show that simple mechanisms can improve steerability.