How do large language models like ChatGPT form opinions about companies?

LLMs do not form opinions in the human sense. They produce a synthesis from two streams: the training corpus they were built on (web pages, news archives, books, Wikipedia, structured datasets) and, at query time, retrieval-augmented generation against the live web. When a user asks about a company, the model assembles an answer from the sources it weights most authoritative on that topic, then renders it in confident prose. The practical implication for a reputation program is that influencing the model means influencing its sources: improving the authority and clarity of what Wikipedia, mainstream news, the company’s own owned properties, and structured data say. Prompting the model directly does nothing. Source-layer work is what moves the answer.