What the Data Actually Supports

The measurement architecture makes it possible to separate a visible change in the buying journey from a claim about faster decision-making. This chapter asks what the passive-log evidence remains strong enough to say after exposure states, journey augmentation, adjustment, and uncertainty are kept distinct. The answer is narrower than the most dramatic interpretation: the visible pre-intent mix changes in a stable way, while broad acceleration and reduced deliberation are not established.

The stable result is a change in visible composition

The strongest result concerns channel mix before intent. Among journeys associated with observed LLM exposure, the adjusted pre-intent share attributed to LLM activity is higher. The reported average marginal effect is +0.009, with a 95% confidence interval from 0.005 to 0.014 and p < 0.001. Review and user-generated-content share is also higher after adjustment, with an average marginal effect of +0.028, a 95% confidence interval from 0.003 to 0.053, and p = 0.027.

These are compositional findings. They describe how the visible pre-intent portion of the journey is distributed across channel classes. They do not mean that an LLM replaced all classic search or that every review disappeared. In fact, classic-search share was not distinguishable from zero after adjustment: the reported average marginal effect was +0.017 with p = 0.292. Retailer or marketplace share and direct-brand share also remained null. The visible mix changed by adding a detectable LLM component and a higher review or UGC component, not by showing a simple collapse of every traditional channel.

The intervals clarify the difference between a measured shift and a dramatic displacement claim. The adjusted LLM-share estimate stays above zero across its reported confidence interval, while the interval for classic-search share includes values compatible with no clear difference. This does not rank one channel as intrinsically better. It says that the study has stronger evidence that the visible LLM component differs than that the visible classic-search component does. The comparison is about what the trace contains, not about the total value of each channel to the consumer.

That distinction matters because channel share has a denominator. It refers to the events that enter the pre-intent trace, not to every act of comparison a person may have performed. If part of a conversation is unobserved, the measured LLM share may still understate the total role of the assistant. Conversely, a rise in visible LLM share does not prove that the assistant displaced an equal amount of classic search. The share tells us how the observed path is composed; it does not recover the unobserved path.

The finding is nevertheless consequential. It means that the measurement environment is not static. An analyst who continues to classify discovery only through search engines, reviews, retailer pages, and brand visits will increasingly describe a different mixture of the journey than the one people are actually exposing to the system. The issue is not that one new channel has been added to a stable funnel. The visible boundary itself is moving.

The accompanying rise in review and UGC share is also important. It complicates a simple substitution narrative in which assistants replace human-written or user-generated evaluation. LLM-exposed journeys in this sample show more visible LLM activity and more visible review or UGC activity after adjustment.