When a customer asks ChatGPT about your company, the answer is not pulled from your website in real time. It is assembled from what the model absorbed during training - Common Crawl web pages, Wikipedia, news, reviews, Reddit threads, job postings and structured databases like Crunchbase and Wikidata - frozen at a knowledge cutoff months before the model shipped. This story explains where that knowledge comes from, why it drifts stale, how gaps and old narratives creep in, and what companies can actually do to correct the record.
A plain-language field guide to how large language models decide what your company is, who it competes with, and whether to mention you at all. AI does not read a homepage the way a customer does. It assembles a mental model from six raw materials - entities, relationships, sources, repetition, consensus, and freshness - and the companies that understand those six levers are the ones that get named when someone asks an AI for a recommendation.