The Journey Did Not Disappear—It Moved
The subject of this series is AI-mediated search and product discovery: what happens when a generated answer or conversational assistant becomes an entry point between a person’s question and the pages, products, or brands they eventually inspect. The immediate question is not whether artificial intelligence is “replacing search.” It is more specific and more consequential: when evaluation moves into an interface that conventional analytics records only partially, what can still be inferred about the journey? The answer matters because marketers may reallocate credit and budget based on the first visible click, even when that click was only the downstream remainder of a longer decision process.
Two recent studies make the problem concrete from different directions. One observes where people look on a search results page that places an AI Overview above traditional ranked links. The other reconstructs consumer journeys from passive behavioral logs and asks how observed LLM exposure changes the visible path toward product intent. Together, they support a bounded conclusion: AI interfaces can change what is visible first and what is counted later. They do not, by themselves, prove that consumers considered fewer options, exerted less effort, or reached a decision faster.
The first visible page is no longer the whole journey
A search-led buying journey is easy to imagine because its traces are familiar. Someone searches for a product, opens review or comparison pages, browses a marketplace, visits several product-detail pages, checks price and shipping, and eventually adds an item to a cart. Each interaction leaves a separate event. A dashboard can arrange those events in sequence and give them names such as discovery, consideration, intent, and conversion. Those labels are useful only as operational descriptions of what the system can recognize. They are not proof that a consumer’s mind moved through the stages in that order, or that the first recorded event was the first meaningful act of discovery.
That distinction becomes important when a business compares channels. A review page may be easy to count because it produces a URL, while a conversational comparison may be represented by one domain or app event. The measurement system then has unequal resolution: it records one kind of evaluation as a chain of visible actions and another kind as a compressed block. Attribution can still produce precise-looking percentages from that data, but precision in the output does not repair a missing part of the path.
An LLM-mediated journey can contain the same underlying questions while exposing a different sequence. In the illustrative example used by the journey study, a shopper looking for noise-canceling headphones under a budget might begin with a prompt that prioritizes comfort. The assistant can produce a shortlist inside the conversation. The shopper may then jump directly to a few product pages, check price and shipping, and proceed toward checkout. The visible log contains fewer review pages and fewer product comparisons, but that does not establish that comparison never happened. It may have been consolidated inside the conversational interface.
This creates an observability boundary. On one side is the person’s evaluation: the alternatives they considered, the criteria they weighed, and the uncertainty they tried to reduce. On the other is the trace available to a search or marketing system: domains, app events, timestamps, clicks, product-page visits, and purchase-adjacent actions. A generated interface can perform several roles at once—information source, comparison surface, recommendation agent, and redirect mechanism—while appearing in a passive log as one touchpoint or as no separately identifiable discovery stage at all.
The consequence is not merely that a channel is missing from a report. A missing early touch can change the strategic interpretation of everything that follows. A product page may appear to begin the journey. A shorter interval between the first visible consideration event and an intent signal may look like improved funnel efficiency. Fewer visible review visits may look like weaker research. In each case, the same downstream behavior can be read as a change in consumer behavior when it may instead reflect a change in where the behavior occurred.
Two instruments, two visibility layers
The evidence in this series comes from two instruments that should be connected carefully, not merged into one experiment.
The eye-tracking study looks inside the search interface. Participants viewed simulated search results with an AI Overview above ten traditionally ranked results while a Tobii device recorded their gaze. This design can answer a focused interface question: which parts of the page receive attention, and how does the generated block alter the path through the results? It cannot observe a complete purchase journey or the private reasoning that follows a glance. The analyzed sample contained 36 participants completing ten tasks in a controlled lab setting. That scale is appropriate for an exploratory view of gaze patterns, but it is not a population estimate of how every searcher behaves or a direct measure of later purchasing.
The passive-log study looks across a consumer journey. It uses de-identified event traces linked to survey data, identifies observed LLM exposure from domain and app signals, and compares journey-level patterns such as visible channel mix, timing, transitions, and breadth. This design can answer a different question: how does the structure of the recorded path differ when LLM exposure is observed? It cannot see the content of the conversation, the depth of comparison inside it, or the effort required to reach a choice. Its final dataset contains 587 users, 3,360 journeys, and more than 2.9 million trace events, with 185 users showing at least one observed LLM touchpoint and 402 showing none. Those counts make the study large enough to examine trace structure, but they do not turn self-selected exposure into a randomized comparison.
The distinction matters because each study sees a different slice of the same broad change. Eye tracking reveals that the generated layer can command attention before traditional links do. Passive logs reveal that early-stage discovery can become less visible in the external path. The first is about attention within an interface; the second is about the trace that remains around an interface. A result from one should not be used as if it directly measured the other.
This separation also prevents a subtler error: treating a well-validated measurement as a complete account of behavior. The passive-log study uses no-lookahead labeling and audits its exposure and stage signals, which strengthens confidence that the recorded categories are reproducible. But a reliable label for an observed touchpoint is still only a label for what entered the log. It does not reveal the words exchanged with an assistant or the alternatives considered without an external click. Likewise, a stable gaze pattern establishes where attention accumulated on the page, not whether the information was accepted as true.
This is why the central issue is measurement, not novelty. If the generated interface absorbs several formerly separate interactions, the analyst needs to know whether a smaller trace represents a smaller journey or a less observable one. That question must be answered before a team treats a lower count of early-stage touches as a performance improvement.
A shorter trace is not necessarily a shorter decision
The word “compression” can conceal two different claims. Behavioral compression would mean that the person actually performed less evaluation or reached a decision with less effort. Observable compression means that the portion of the journey recorded by the measurement system became shorter because some evaluation moved into a less visible environment. The passive-log study is designed to measure the second claim. Its authors explicitly treat mechanisms such as shortlist formation and consolidated comparison as interpretations consistent with the observed patterns, not as directly identified processes.
That boundary also changes how attribution should be read. Traditional path models use order and adjacency as clues about channel roles. If the first observable non-LLM touch is a retailer product page, the model may assign that page more influence than it deserves because the preceding conversational discovery is absent. A dashboard may then reward the visible remainder and encourage more investment in tactics that merely appear to start the journey. The problem is not solved by making the model more elaborate if the relevant touchpoint is still outside the measured path.
The same caution applies to apparent timing improvements. A direct jump from consideration to a small set of product pages can look like faster movement toward intent. But a direct jump may mean that comparison was front-loaded into the assistant, not that the person stopped comparing. The visible sequence has changed; the underlying amount of deliberation remains unmeasured. Absence of an early event is therefore not evidence that no early evaluation occurred.
This does not make the data useless. It changes the question the data can answer. A dashboard can still detect that the visible mix of channels is changing, that a new touchpoint is appearing earlier, or that direct transitions are becoming more common. Those are important signals about the measurement environment. What the dashboard cannot responsibly do is silently translate those signals into a claim about reduced cognitive effort or universal improvement in conversion efficiency.
The practical judgment this series supports is consequently narrow but useful. AI-mediated discovery should be treated as a change in the measurement environment. Analysts need to ask whether an AI interface has become a first touch, whether its activity is recorded as a distinct channel, and whether a shift toward direct downstream transitions is a performance signal or an observability alarm. The next step is to examine the interface itself: if generated content receives the first and strongest attention, how does that alter the role of the traditional results that remain below it?