The Journey Is Not the Script: Reconciling Structured and Naturalistic Probe Design in Multi-Turn Brand Recommendation Audits

Recent work on conversational information seeking argues that a single prompt, whether the first or the last, is a poor proxy for the full request state of a multi-turn AI conversation (Tannenbaum, 2026). Across 8,133 real human-LLM conversations, that study finds that the final user turn carries a median of roughly one-third of a session's content vocabulary, that request-state dimensions established earlier are frequently left unrestated at the endpoint, and that new criteria are introduced late, including at the very last turn, in a material share of sessions. This raises a fair and specific question for any commercial audit methodology that measures brand recommendation across multi-turn AI conversations: does a structured probe script, one that assumes a conversation moves predictably through fixed stages, miss the kind of late, path-dependent state changes that occur in real conversations? This note describes the two probe architectures used in AIVO Meridian audits, a fixed-stage structured probe and a longer, less constrained agentic probe, and argues that the two are designed to answer different, complementary questions rather than the same one. It states plainly where the structured design's assumptions are vulnerable to exactly the critique above, and proposes the direct empirical comparison needed to test whether the agentic probe's displacement measurements capture the late-turn state changes Tannenbaum documents. That comparison has not yet been run. We treat it as the necessary next study, not a settled result.