I study how linguistic features of corporate disclosure shape earnings expectations. I develop a methodology that uses large language models to generate counterfactual versions of earnings call transcripts that differ in language while holding underlying quantitative content fixed, creating a form of counterfactual variation that is impossible to obtain experimentally. I first show that textual features from earnings calls improve out-of-sample predictions of analyst forecasts and realized earnings above and beyond a rich set of quantitative fundamentals. I then intensify each transcript along six linguistic features---Forward Guidance, Jargon, Macro Focus, Uncertainty, Sentiment, and Confidence---and trace how these shifts move predicted analyst expectations. The results show that analysts move their forecasts in the right direction but weight linguistic dimensions unevenly: they place excessive weight on optimistic and macro-focused language and do not fully incorporate information embedded in language emphasizing risks and uncertainty. Linguistic shifts also change which fundamentals drive predictions: optimistic language amplifies the role of recent earnings surprises, while risk-laden language diminishes it, evidence that presentation shapes not just the level of analyst beliefs but the information they attend to.