I investigate whether managers learn from conversations about their firm on Twitter when forming their expectations about future firm performance. I conjecture that tweets addressed to or about the firm, as initiated by its different stakeholder groups, such as customers, employees, or suppliers, contain useful and timely forward-looking information. I use a machine-learning-based natural language processing algorithm to analyze 11 million stakeholder-initiated tweets before the announcement of 4,131 sales forecast from 2013 to 2020. I find evidence that managers could learn from listening to conversations about their firm on Twitter and improve forecast precision before they announce their forecasts: I show that information in conversations is negatively associated with the directional error in management sales forecasts, suggesting that managers do not adequately account for third-party information: When the conversational information is negative, managers overestimate future sales. In contrast, managers tend to underestimate future sales when it is positive. Further, I find the relevance of Twitter information increases the closer the information is released to the forecast announcement date.