Repeated Queries Exhaust an LLM's Brand Recommendations but Not Its Sources

Research on 300 combinations of question-answer engines (50 questions, 6 engines, 15 runs) found that repetitive purchase questions would exhaust the brand recommendation list of LLMs enabled for online retrieval, but not their knowledge sources. In 86-92% of non-indexable engine cells, brands never seen before were still being added by the 15th run, with a median of 15-31 organizations; while the list of engines enabled for retrieval saturated by the 10th run (median 8 organizations). Additionally, the accumulation of reference domains continued to rise at each testing point: four deep cells used domains in the 24th run, with 59-84% of Chao2 lower bounds observed; 44% of breadth cells of indexable engines were still adding domains by the 15th run. A single run covered only 62-77% of the brand set over five runs, with a median of 38 organizations extracted from engines, and the median was 15 organizations that appeared in only one engine.