The Core Question
Can the movement of a mouse cursor and the rhythm of a scroll reveal where human attention actually goes? The intuition is plausible: people tend to move their cursor toward content they are reading, scroll at speeds that reflect their reading pace, and pause when something arrests their attention. But intuition is not evidence. The scientific literature on this question now spans over two decades and multiple independent research groups. The conclusion is consistent: cursor and scroll dynamics are a valid, measurable proxy for visual attention, validated against eye-tracking as the gold standard.
The Cambridge Validation: MouseView.js
The pivotal peer-reviewed validation comes from Anwyl-Irvine et al. (2021), conducted at the MRC Cognition and Brain Sciences Unit, University of Cambridge. The study validated MouseView.js, an open-source JavaScript tool that measures mouse-contingent viewing, directly against established eye-tracking methodology in controlled conditions.
The finding was unambiguous: MouseView.js proved equally reliable as eye-tracking and produced the same pattern of dwell-time results across both the primary study (N=165) and a free-viewing replication (N=83). The dwell-time differences measured by MouseView.js and by eye-tracking were highly correlated and related to self-report measures in comparable ways.
Three aspects of this finding matter for practical application. First, the correlation holds at the level of behavioral patterns, not just directional trends. Second, the tool is implemented in standard JavaScript, requiring no specialist hardware and deployable at census scale through existing web infrastructure. Third, the study was conducted by an independent academic institution with no commercial stake in the outcome.
Independent Replication Across Domains
Scientific findings gain authority through independent replication. The Anwyl-Irvine validation has been replicated across multiple independent research contexts since publication. Woronko et al. (Journal of Anxiety Disorders, 2023) used MouseView.js methodology to study attentional bias in specific phobia, confirming its validity as an attention measurement instrument in a clinical research context. Milani et al. (Journal of Sex Research, 2025) applied the methodology to motivational stimulus processing, again confirming its reliability as an attention proxy.
These replications are significant because they demonstrate that the cursor-attention correlation holds across different content types, different emotional valences, and different research populations. The validity of cursor tracking as an attention proxy is not limited to neutral reading tasks; it extends to the full range of human attentional engagement with digital content.
The Research Lineage: Two Decades of Evidence
The Cambridge validation builds on a research lineage extending to the early 2000s. Johnson et al. (2012) demonstrated that mouse movements could effectively track visual attention in advertising contexts. Arapakis et al. (2014), published in ACM SIGIR, the primary academic venue for information retrieval research, examined how participants interacted with online news articles using mouse tracking as an engagement proxy. Huang et al. (2012) investigated the relationship between eye position and cursor position during web search, identifying distinct behavioral modes: reading, hesitation, scrolling, and clicking, each with characteristic cursor signatures.
Arapakis and Leiva's subsequent work at ACM SIGIR (2016, 2020) extended this lineage to machine-learning prediction of user engagement from cursor data, and to learning efficient representations of mouse movement sequences for attention prediction. The consistency of findings across over a decade of independent research, published in peer-reviewed venues ranging from cognitive neuroscience journals to computer science conferences, constitutes a robust scientific foundation.
Scroll Depth and Attention Decay: The Ebbinghaus Connection
Cursor tracking addresses the question of where attention goes within a content piece. Scroll depth and velocity address the question of how far attention penetrates into content and how it decays. The scientific foundation here connects to one of psychology's most replicated findings: the Ebbinghaus forgetting curve of 1885.
Ebbinghaus established that memory and attention follow exponential decay functions, a finding replicated consistently across 150 years of cognitive psychology research. Applied to digital content, the same decay pattern emerges empirically. Research on online content consumption shows that peak viewership occurs within 5 days of content publication in 73% of documented cases, followed by exponential decay driven by novelty depletion and competition for attention from newer content.
Segmented regression models applied to normalized daily pageview data on logarithmic scales confirm that exponential decay fits outperform power-law models in the majority of cases. This is the mathematical pattern that exponential decay models apply to individual content pieces: not an assumption imported from laboratory psychology, but an empirical regularity confirmed in production content environments.
Reader Behavioral Segmentation: From Six Types to Seven
The scientific literature on reader behavioral typologies provides direct empirical grounding for cognitive segmentation approaches in attention measurement. From scroll interaction data alone, researchers have identified six distinct reading behaviors with characteristic metrical signatures: bounce back, characterized by immediate exit; shallow, characterized by minimal scroll depth; scan, characterized by rapid scroll velocity; idle, characterized by dwell without scroll; read, characterized by moderate depth and velocity; and read long, characterized by extended dwell and full depth penetration.
A more granular classification comes from eye-tracking research using within-subject experimental design with 30 participants across six simulated browsing tasks, with scanpath data classified by three independent machine learning approaches including LSTM, random forest, and multilayer perceptron classifiers, published in User Modeling and User-Adapted Interaction in 2024. This confirms that behavioral signals can reliably distinguish between attentive, comparative, reading, and free-browsing states, even without direct eye-tracking hardware, when the behavioral signal collection is sufficiently granular.
At the apex of these typologies sits Deep Reading, defined in the cognitive science literature as the full sum of cognitive, perceptual, and affective processes that prepare readers for apprehending, understanding, and absorbing the gist of what is read. Predictive models of deep reading comprehension using gaze-derived features have been validated on 247 participants across two independent studies, establishing the connection between observable behavioral signals and the cognitive state that content producers most want to reach.
Why Census Scale Changes Everything
The scientific validation of cursor and scroll tracking as attention proxies has a practical implication that panel-based eye-tracking cannot match: census-scale deployment. Eye-tracking panels measure real attention from real eyes, but from samples of tens or hundreds of participants recruited and equipped with specialist hardware. The statistical extrapolation from panel to population introduces uncertainty that grows as content environments diversify.
Cursor and scroll tracking, implemented through standard JavaScript on existing web infrastructure, measures attention signals from every actual visitor to every actual piece of content. The measurement population is the actual audience, not a sample of it. This is not a marginal methodological difference; it is a structural advantage that becomes more significant as content environments fragment and audience segments narrow.
The scientific case for cursor-based attention measurement is not that it perfectly replicates eye-tracking. It is that it measures the same underlying phenomenon, validated against the same gold standard, at a scale that eye-tracking cannot approach, without specialist hardware, and with a privacy architecture compatible with current regulatory requirements.