Linguistic features of a message necessarily shape its persuasive appeal. However, studies have largely examined the effect of linguistic features on persuasion in isolation and do not incorporate properties of language that are often involved in real-world persuasion. As such, little is known about the key verbal dimensions of persuasion or the relative impact of linguistic features on a message’s persuasive appeal in real-world social interactions. We collected large-scale data of online social interactions from a social media website in which users engage in debates in an attempt to change each other’s views on any topic. Messages that successfully changed a user’s views are explicitly marked by the user themselves. We simultaneously examined linguistic features that have been previously linked with message persuasiveness between persuasive and non-persuasive messages. Linguistic features that drive persuasion fell along three central dimensions: structural complexity, negative emotionality, and positive emotionality. Word count, lexical diversity, reading difficulty, analytical language, and self-references emerged as most essential to a message’s persuasive appeal: messages that were longer, more analytic, less anecdotal, more difficult to read, and less lexically varied had significantly greater odds of being persuasive. These results provide a more parsimonious understanding of the social psychological pathways to persuasion as it operates in the real world through verbal behavior. Our results inform theories that address the role of language in persuasion, and provide insight into effective persuasion in digital environments.