LLM exposure and funnel erosion in consumer journeys: measuring journey compression and reordering in passive behavioral logs
Journal Article
Overview
abstract
Abstract; LLM assistants increasingly mediate product discovery, yet most journey analytics and attribution systems still assume that discovery begins with search, retailer browsing, or direct brand visits. That assumption misstates the observed journey. When evaluation moves into conversational interfaces, passive behavioral logs undercount early-stage activity and standard journey reconstruction can resemble funnel erosion. We develop a measurement framework that treats LLM usage as a first-class touchpoint in de-identified passive event logs linked to a post-trace survey. The framework detects observed LLM exposure from domain and app signals, operationalizes a non-linear stage codebook (consideration, intent, conversion-proximate), and computes journey metrics for compression, timing, channel mix, and transition structure without imposing a canonical path. A production-scale, no-lookahead labeling system combines deterministic cues, model-assisted classification, locked holdouts, and targeted adjudication audits to deliver reproducible measurement in noisy passive logs. In a one-month observation window, the most stable empirical result is a shift in visible pre-intent discovery: LLM-exposed journeys show substantially higher observable LLM share and higher review/UGC share, while unadjusted compression and time-to-intent differences shrink after adjustment. Conversational discovery changes what marketing analytics systems observe, and journey measurement must be updated accordingly.
