What the Eye Sees First
This chapter examines a specific part of AI-mediated search: what changes when a generated overview occupies the top of a search results page before the traditionally ranked links. The eye-tracking evidence helps answer a practical question for search and marketing teams: does the new block merely add information, or does it reorganize the first moments of attention? The study shows a meaningful reallocation of visual attention, while also showing that the familiar search pattern has not vanished. That combination is important because attention can move upward without traditional search becoming irrelevant.
A controlled view of the new search page
The study used a within-subjects design. The same participants completed all ten search tasks, each built around a short backstory and a single simulated search results page. A generated AI Overview appeared above ten Traditionally Ranked Results. Participants could inspect the page and open results until they felt their information need had been met, while the study recorded gaze and interaction behavior. The analyzed sample contained 36 participants.
This design gives the comparison a useful internal reference. The question is not whether people look at an AI Overview in a vacuum; it is how attention is distributed when the overview and the familiar ranked list occupy the same page. Because the same people completed the task set, differences between the two regions are less dependent on comparing entirely separate audiences. The design still remains a controlled experiment rather than a record of ordinary search across months of changing needs. Queries were pre-typed, the interface was simulated, and participants could not begin a new search session or refine a query after starting a task.
The study also separated the page into Areas of Interest: the AI Overview and the traditionally ranked results. That matters because “attention” is not one measurement. Fixation time captures how long the eyes remain relatively still in a region; saccade time describes the rapid movement between fixations; saccade length records the distance of that movement. These measures are clues about visual examination and navigation, not direct readings of comprehension or belief. A person can look at a summary carefully, doubt it, and then search for corroboration. Conversely, a brief look can be sufficient for a familiar or simple task.
The study therefore answers a narrow interface question. It can show which region draws gaze and how that distribution changes when generated content is placed first. It cannot show how much comparison occurred inside an AI system, whether participants accepted the summary as true, or whether the page changed a later product decision. That distinction is the eye-level version of the wider observability problem: the measurement is valuable precisely because it is specific.
The top of the page wins attention
Across the ten search results pages, participants spent an average of 45.73 seconds viewing the SERP. The AI Overview received an average of 27.34 seconds of fixation time, compared with 14.82 seconds for the traditionally ranked results. Mean saccade time was 3.15 seconds for the AI Overview and 1.85 seconds for the ranked results. The reported differences in fixation and saccade time were statistically significant; the difference in saccade length was not.
The simplest interpretation is not that participants abandoned the ranked list. It is that the generated block became the main early object of inspection. It occupies the top of the page, offers a synthesized answer, and can present a coherent response before the searcher has opened any external result. In the old list-only layout, the first ranked result was the most prominent candidate for that first sustained look. With an AI Overview in front, the first sustained look is redirected to a new kind of object.
That change has a direct implication for how visibility should be understood. A traditional result can be relevant and still receive less initial attention because the page now offers a summary that appears to address the need before the list begins. The generated block is not merely another link competing for a click. It is an answer-shaped region that can collect attention without requiring a visit to the source pages represented beneath it. If the searcher later clicks a ranked result, that click occurs after the overview has already influenced the sequence of inspection.
The important unit here is not the isolated gaze event but the order of exposure. A searcher first encounters a synthesized response, then decides whether the response is sufficient, questionable, or incomplete, and only then decides how much of the ranked list to inspect. The ranked result therefore competes against an interpretation that has already been formed on the page. A traditional result can still win that competition, but it no longer begins from the same informational starting line.
That has implications for both search visibility and measurement. A link may receive fewer early fixations because the overview has answered the broad question, while still becoming important when the searcher needs evidence, detail, or a source to verify. Conversely, a summary can receive heavy attention because it is easy to scan, not because the searcher has accepted every statement in it. The metric tells an analyst where the interface concentrates the opportunity to influence evaluation. It does not settle the later judgment about accuracy or usefulness.
The within-subject design helps isolate this page-level redistribution, but it does not remove the distinction between exposure and effect. Participants saw the same broad task set, yet their experience was still shaped by the controlled environment, the prepared queries, and the particular content rendered in the overview. The result is strongest as a statement about what the layout makes more visible under the tested conditions. It becomes weaker when converted into a claim about all topics, all users, or all forms of generative search.
The mechanism can be stated simply:
This sequence does not imply that placement alone determines what people believe or choose. The study’s own results point to other influences, including task complexity and the content’s perceived usefulness. But placement changes the order in which evidence becomes available. An answer that appears first can shape what the searcher treats as a sufficient starting point, what they decide to verify, and which traditional result becomes the first external reference.
For marketers, the practical point is not that a generated block automatically wins the customer. It is that the first visible opportunity to influence evaluation may no longer be the first ranked link. A click-only report can miss that early competition because the AI Overview can receive substantial gaze without producing a conventional outbound event. Attention is therefore a leading indicator of interface exposure, not a substitute for a conversion measure.
Why normalization changes the interpretation
There is an obvious alternative explanation for the raw fixation difference: the two page regions do not have the same physical or textual size. Across the ten pages, the traditionally ranked-results region was about 2.7 times larger in pixel area and contained about 2.8 times more words than the AI Overview. A larger region offers more material to inspect and more opportunities for fixation. Comparing only total seconds would therefore risk attributing to the generated format what might partly be an artifact of layout.
The researchers addressed this by calculating fixation time normalized by the total word count of each region. On that measure, the AI Overview averaged 170.39 milliseconds per word-count unit, compared with 33.99 for the traditional-results region. The normalization does not turn the result into a universal measure of reading speed, and it does not erase every difference between an answer and a list. It does, however, make the interpretation more demanding: the overview attracted more attention relative to its textual size, not simply more attention because the page contained more words below it.
Normalization is useful here because interface comparisons often create unequal objects. An AI summary may be shorter but denser; a ranked list may be longer but distributed across repeated snippets and links. If analysts compare only raw exposure time, they may confuse the amount of content with the way the content is examined. If they normalize too aggressively, they may hide meaningful design differences. The right conclusion is more modest: both the raw and normalized comparisons indicate that the generated region received unusually strong visual attention in this controlled setting, while the metric still describes attention to a page region rather than understanding of its claims.
This is also why the first ranked result should not be judged by raw attention alone. Its exposure is conditioned by the presence, length, and apparent completeness of the overview above it. A lower fixation share could mean that the overview met the need, that the searcher postponed verification, or that the page’s visual hierarchy discouraged further scanning. The eye data cannot choose among those explanations by itself. It identifies a changed opportunity structure that other measures—link selection, follow-on visits, self-reported relevance, and longer-run journey traces—would need to interpret.
For a marketing team, the distinction changes what counts as a meaningful visibility question. The question is no longer only “Which ranked result received the first click?” It is also “Which part of the page framed the decision before that click?” That framing can affect the probability that a searcher reaches a brand page, the time they spend deciding whether to continue, and the role of a source link as confirmation rather than discovery. None of these outcomes is guaranteed by the gaze result, but each becomes harder to analyze if the generated region is treated as empty space above the real search results.
The same caution applies to pupil dilation and saccade behavior. The study reports several physiological measures and explores associations with task ratings, but a physiological signal should not be treated as a direct scale of confidence, effort, or purchase intent. The value of the eye-tracking result lies in the converging page-level pattern: fixation and saccade time favor the overview, and normalized attention does not make that difference disappear.
The golden triangle moved upward
The classic “golden triangle” or F-shaped scan pattern remains visible in the aggregated heatmaps. Searchers still concentrate on the top-left portion of the results page, and attention still declines through much of the ranked list. What changes is the content occupying the most valuable location. The top-left region now contains an AI Overview before it contains the first traditionally ranked result.
The comparison of fixation proportions makes the shift especially clear. In the AI Overview condition, the overview received 0.59 of total fixation duration, while the first ranked result received 0.09. In a traditional-results-only comparison, the first ranked result received 0.31. The study interprets this as an effective displacement: the first ranked link remains first among the traditional results, but it no longer occupies the page’s first opportunity for sustained attention.
“Displacement” should be understood as a statement about attention allocation, not a claim that the ranked result moved or became less relevant. Its position in the list did not change. Its competitive context did. A result that once benefited from being the first substantive answer now appears after an answer-shaped interface element. Its performance is consequently influenced by whether the overview satisfies the searcher, raises a question, creates doubt, or sends them looking for corroboration.
This is why the persistence of traditional scanning matters. The generated layer changes the starting condition, but the ranked list retains its own order, link structure, and declining pattern of attention. A searcher can read the overview, continue downward, and use a traditional page to confirm or challenge it. The page is not divided into “AI” and “search” as mutually exclusive worlds. It is a single sequence in which one layer changes the visibility and timing of the other.
The sequence also explains why traditional results can remain strategically important even when their first-glance share falls. A source page may be less likely to receive an exploratory visit but more likely to receive a deliberate verification visit. Those are different roles, and a metric that groups them together will miss the change. The study does not measure that role transition directly, but its attention pattern makes the possibility visible: the generated layer captures the broad first look, while the ranked list remains the place where a searcher can pursue independent detail.
What gaze can and cannot tell us
The eye-tracking findings establish a strong interface observation: when an AI Overview appears above traditional results, it receives more visual attention than the ranked-result region, and the top-left scan pattern persists around that new anchor. They do not establish that the overview is more trusted, more accurate, or more useful for every task. They do not show that people stop visiting external pages, and they do not measure the hidden comparison that may occur in a conversational system before or after the page view.
The limits are substantive. The study is exploratory and lab-based, uses a simulated interface and pre-typed queries, and does not allow later search sessions or query refinement. The task-complexity groups and overview sizes are not perfectly balanced, and the sample is small. These conditions make the result best suited to explaining a mechanism of attention, not forecasting all search behavior.
That mechanism is still consequential. If the first visual encounter changes, then the boundary between “seen,” “considered,” and “clicked” changes too. The AI Overview can become the first place where a searcher forms a working interpretation, while traditional links remain available for detail or verification. The next question is therefore not whether the ranked results vanished. It is what people do with them after the generated answer has already shaped the first glance.