I train machine learning models on earnings call text and fundamentals to predict analyst expectations and realized earnings. Using LLMs, I generate counterfactual transcripts that amplify specific narratives. Comparing model predictions for these morphed transcripts to those for the real earnings calls allows me to estimate the narrative predicted treatment effect on analyst beliefs and future earnings. The results reveal systematic biases: analysts over-react to sentiment (optimism) and under-react to narratives of uncertainty.