Decision-Path Analysis across AI Recommendation Systems

Abstract This paper presents the first evidence-grade taxonomy of decision filters applied by large language models (LLMs) during AI-assisted consumer purchase journeys. Drawing on 12 months of live multi-turn transcript analysis across ChatGPT, Gemini, Perplexity, and Grok, covering 160+ brands across six product categories and 7,000+ four-turn buying sequences, we identify eight structurally distinct filter types that determine recommendation outcomes at the decision stage of AI buying conversations. We distinguish the current analysis from prior AI visibility research in three ways. First, findings are derived from verbatim transcript data at the turn level — not from win/loss outcome classification alone. Second, the methodology is multi-turn, capturing the full decision-path from awareness through to purchase recommendation (T1 through T4). Third, filter identification is evidence-grade: we quote the AI's own dismissal and selection language as primary evidence rather than inferring filter mechanisms from aggregate outcome data. Key findings include: the systematic elimination of botanical-heritage skincare brands by a Clinical Evidence Binary filter operating at T3; a cascade displacement phenomenon in which premium brands lose to value alternatives on a value-per-clinical-evidence axis; a Close Second Trap pattern in banking in which brands acknowledged as strong runners-up are never recommended; and a Technology Generation Tiebreaker producing model-specific divergence in haircare. We propose a Filter Taxonomy of eight types and discuss implications for brand positioning strategy, remediation methodology, and AI search measurement practice.