We Described One Brand to the Model, Real or Made Up. It Recommended That Brand Anyway.

This study examines whether clearer framing of a brand's own facts changes whether GPT-4o recommends that brand. The same underlying facts were presented in three ways: clear, ambiguous, and misaligned framing. The study progressed through five experimental designs covering 13 brands, including 9 well-known brands, 3 genuinely obscure real brands, and 1 fully invented brand. Across the five designs, the study reports 1,997 real GPT-4o API calls. For 12 of 13 tested brands, framing quality did not materially change candidacy. The largest outlier showed a 70 percentage-point selection swing between framing conditions. The most important diagnostic was the invented-brand test. A fully invented brand with no possible prior history reached 100% candidacy and 100% selection when its facts were placed in the system prompt immediately before a matching-category buyer question. The result suggests that, under this specific experimental setup, presence in context can dominate the quality of framing itself. It does not establish that framing never matters. The experiment uses GPT-4o and controlled prompt conditions. The study does not claim that these results generalize to all models or conversational structures.