What Makes a Good Query? Measuring the Impact of Human-Confusing Linguistic Features on LLM Performance