Non-Retrieval LLMs Never Stop Naming New Brands: 86-92% of Cells Still Adding at Run 15

arXiv 2609.05059·medium signalXBlueskyLinkedInCopy link

arXiv 2609.05059 tests whether repeated identical buying questions exhaust an LLM's brand recommendations across 300 question-engine cells (50 questions, six engines, 15 runs each, 1,470 adjudicated organizations). The five engines answering without web search were still adding never-seen brands at run 15 in 86-92% of cells with median repertoires of 15-31 organizations, while the one retrieval-enabled engine closed its list at a median of 8. A single run shows only 62-77% of the five-run brand set, and the paper shows fixed-roster extraction manufactures plateaus that open extraction removes — a direct warning for anyone measuring brand visibility in LLM answers.