Embedding as a pitfall for survey‐based welfare indicators: evidence from an experiment