Enterprises still assume visibility inside AI assistants can be managed with search-era tools such as content optimisation, metadata, schema, and authoritative signals. This article shows why such assumptions fail. Large Language Models reconstruct rather than retrieve, producing unstable, contradictory, and fabricated representations of companies even when inputs and disclosures remain unchanged.