Overview

AI-mediated search measurement needs separate states for user-level observed LLM exposure and journey-level pre-intent augmentation, then a non-linear stage codebook for consideration, intent, and conversion-proximate behavior. This chapter shows how compression, timing, channel mix, transitions, and source breadth answer different visibility questions, and why no single metric recovers hidden effort. It also distinguishes label validation from causal validity: audited exposure and stage labels can be reproducible without making self-selected groups comparable or revealing conversation content. The result is an architecture that preserves unit, scope, and uncertainty before adjusted associations are interpreted.

Learning outcomes

Observed LLM exposure belongs to the user and window; pre-intent augmentation belongs to a particular journey and cannot be inferred for every path.
Compression, timing, channel mix, transitions, and breadth describe different parts of the visible trace and should not be collapsed into one score.
Validation can strengthen confidence in labels without proving that an LLM touch caused the journey pattern.

Keywords

LLMExposureAugmentationMetricsValidationCausality

Measure the Interface Before You Measure the Funnel

The central measurement problem in an AI-mediated buying journey is not a lack of metrics. It is a failure to keep different kinds of state separate. This chapter examines the architecture needed to tell whether a person was exposed to an LLM, whether a particular journey was augmented by one before intent, and how the visible events in that journey should be staged. The distinction matters because a well-labeled event can still be an incomplete view of the decision, and a precise association can still be non-causal.

Exposure is not the same as augmentation

The journey study begins with a broad user-level exposure registry. Let observed exposure mean that a user has at least one detected LLM touchpoint during the observation window. This state belongs to the person and the window. It answers a descriptive question: did the measurement system see this user encounter an AI or LLM surface at all?

That state is deliberately broader than the journey-level augmentation measure. A particular journey is marked as pre-intent augmented only when it contains a stricter canonical LLM touchpoint before the first observed intent event. This state belongs to the journey, not automatically to every journey made by the same person. A user can have one journey that includes a pre-intent LLM touchpoint and another that does not. The framework enforces that a journey cannot be marked augmented if the user has no observed exposure at all, but the reverse is allowed: exposure does not imply augmentation in every path.

The distinction prevents a common analytical shortcut. Suppose a person uses an assistant while researching one product and later visits a retailer directly for another. If every journey from that person is labeled “AI-augmented,” the analysis confuses a person-level propensity with a journey-level mechanism. It may then attribute differences to the interface when they actually reflect who uses assistants, how active they are online, or what kinds of needs brought them into the sample.

The study makes the separation visible through three groups. G0 contains non-exposed users and their journeys. G1 contains exposed users whose particular journeys are not augmented before intent. G2 contains exposed users whose journeys are augmented before intent. G1 is important because it shows that observed exposure and journey augmentation are not interchangeable. Without it, analysts would have no way to tell whether a difference belongs to LLM-exposed people generally or to the journeys in which the LLM touchpoint appears at the relevant stage.

The final dataset illustrates the scale and the asymmetry of the comparison. It contains 587 users, 3,360 journeys, and 2,917,655 trace events. There are 185 users with at least one observed LLM touchpoint and 402 without. At the journey level, the decomposition includes 2,258 G0 journeys, 890 G1 journeys, and 212 G2 journeys.

Published by DeepMox

FAQ

What do G0, G1, and G2 represent?

G0 contains non-exposed journeys, G1 exposed but non-augmented journeys, and G2 exposed journeys with pre-intent augmentation. G1 separates user exposure from the journey-level mechanism.

Are compression and timing the same metric?

No. Compression describes the recorded share of later-stage events, while timing measures the visible interval between first consideration and first intent; neither alone reveals hidden effort.

Does accurate LLM labeling prove causality?

No. It supports reproducibility of operational labels, but it does not establish exchangeability, reveal conversation content, or show that an LLM touch caused the journey pattern.

Previous8 minAI-Mediated Search Buying Journeys

Next10 minAI-Mediated Search Evidence

AI-Mediated Search Measurement Architecture

Measure the Interface Before You Measure the Funnel

Exposure is not the same as augmentation

FAQ