TR-2026-001: How AI Assistants Select Brands to Cite

Report TR-2026-001 breaks the route from a user's question to a shortlist of named brands into four steps: retrieval, candidate assembly, verification, and rendering. Each step removes brands for its own reason, and being findable in ordinary search only clears the first one. The episode also covers two recurring failure patterns and why one test prompt tells you nothing. Retrieval — long questions get split into narrow ones, so pages built around a single narrow topic get fetched Candidate assembly — a name in a list or a logo with no text around it leaves nothing to extract Verification — a spec only the vendor states counts for less than one repeated by outside sources Rendering — limited answer length favours brands whose category takes a few words to state Two failure patterns: brands widely mentioned but never described, and brands only they describe Volatility — repeat answers vary, so position needs repeated observation across phrasings Full report: TR-2026-001: How AI Assistants Select Brands to Cite Trellner Research — trellner.com