Do AI assistants hold consistent brand preferences? Wording effects and a conversational unit in AI B2B software recommendations
- Question
- When you rephrase a buying question, the brands an AI recommends mostly change. Does a stable opinion exist underneath the wording noise, and does asking as a short conversation reveal it better than asking once?
- Status
- Data collected, analysed and frozen after four audits. The write-up goes up after the registration embargo ends on October 26, whichever way the result came out.
Twenty B2B software buying situations, each asked five ways (category-first, a cosmetic rewrite, problem-first with no category named, worth-it, and one added constraint), of ChatGPT, Gemini, and Claude with each assistant's own web search on, three reruns each. Three ways of asking: a single question, a single question that forces one pick, and a two-turn conversation that asks then forces the pick. The outcome is a reliability coefficient (generalizability theory) on which brand gets picked, split by whether the model searched or answered from memory, reported by brand prominence tier, with a positive control where a stable pick plainly exists and a null control of two identical wordings to bound noise. Three AI extractors plus 400 hand-coded answers with an agreement floor. Follows the last line of Jack et al. (2605.27440), that better measurement needs a different unit, and Żatuchin's variance-components method (2607.13304). Preregistered, two-sided, hard cost cap, B2B software only. What it finds may inform the product.