Research / GPL-TR-2026-03

Attention Beyond Activity

The gap between action and attention.

When is recorded activity evidence of attention?

Author
GrayPass
Version
GPL-TR-2026-03
Date
2 August 2026
Read the full report

Finding

Activity is not attention.

Activity is not a self-validating measure of attention.

Dashboards report time-on-page, cursor movement, and clicks in the vocabulary of attention. Across 2,775 search sessions where participants also said whether they attended, the standard proxies disagree with each other, half of recorded dwell shows no observed focused activity, no proxy meaningfully tracks the person's own report, and every headline metric can be inflated arithmetically without adding evidence about anyone. The remedy is not to abandon logs but to say what they measure.

Visual analysis

01 / Construct

A log records exposure and activity, not attention itself.

Moving from page exposure to a claim about attention requires several additional assumptions. The ladder makes each inference visible.

FIG. 1The construct ladder

Measurement model · no data plotted

Each step upward requires an additional inference. Comprehension, memory, and intent sit above this ladder and are not measured by the interaction proxies in the report.

02 / Disagreement

The proxies disagree with one another.

Common engagement measures rank the same sessions differently. Longer dwell often means a smaller active share.

FIG. 2Twelve proxies do not tell one story

Attentive Cursor Dataset · 2,775 sessions

0.218Attentive Cursor Dataset · 2,775 sessions

Median absolute correlation among all proxy pairs

-0.835Attentive Cursor Dataset · 2,775 sessions

Dwell versus active fraction, Spearman rho

81.3%Attentive Cursor Dataset · 2,775 sessions

Sessions in discordant median-split quadrants

Dwell versus active fraction

81.3% discordantAttentive Cursor Dataset · 2,775 sessions
Discordant quadrantsConcordant quadrants
Bars show selected Spearman rank correlations among proxies computed on the same retained sessions. Dwell and active fraction move strongly in opposite directions, while several familiar proxy pairs agree only modestly.

03 / Recorded time

A page can stay open after focused activity stops.

Focus, blur, and event gaps separate raw dwell from the portion accompanied by observed activity.

FIG. 3Recorded dwell exceeds observed focused activity

Attentive Cursor Dataset · 2,775 sessions

10th
15.3%
25th
31.5%
Median
56.8%
75th
90.1%
90th
100.0%

Active share of raw dwell · 2 s idle threshold

51.4%Attentive Cursor Dataset · 2,775 sessions

Median dwell idle or blurred

66.1%Attentive Cursor Dataset · 2,775 sessions

Sessions with at least 30% unattributable dwell

63.0 daysAttentive Cursor Dataset · 2,775 sessions

Longest recorded dwell

97.4%Attentive Cursor Dataset · 2,775 sessions

Raw corpus dwell held by that one page

The active-to-raw ratio uses the dataset's own focus, blur, and event-gap structure with the report's primary two-second idle threshold. A single left-open page accounts for most raw corpus dwell, demonstrating why raw sums need qualification.

04 / Criterion validity

Activity barely tracks the participant's own report.

No single proxy has a strong relationship with self-reported attention. A combined model remains only modestly above chance.

FIG. 4Every proxy versus self-reported attention

Attentive Cursor Dataset · 2,775 sessions

-0.10-0.050.00+0.05+0.10

Spearman ρ with 95% intervals · 2,775 retained sessions

0.558Attentive Cursor Dataset · 2,775 sessions

All-proxy model AUC for high self-reported attention

0.792Attentive Cursor Dataset · 2,775 sessions

The same model's AUC for the mechanical click

Points are Spearman correlations against the participant's self-reported attention response. Lines show 95% confidence intervals. Associations remain weak, and several intervals cross zero.

05 / Stress test

Headline metrics can move while the person does not.

Deterministic transformations inflate dwell, event counts, and clicks without adding attention-relevant information.

FIG. 5The metric moves without new attention evidence

Attentive Cursor logs · frozen transformation

Mousemove densification

Insert duplicated points while retaining the recorded path.

Mousemove count

58.75×Attentive Cursor logs · frozen transformation
1× original60× scale
58.75×Attentive Cursor logs · frozen transformation

Mousemove count after the selected transformation

1.00×Attentive Cursor logs · frozen transformation

Median path-length inflation in the duplication variant

Each stress test is a deterministic transformation of frozen logs. It changes metric arithmetic, not human behavior. The controls show a related quantity that remains bounded or unchanged.
Interaction logs can describe what they record. Claims about attention require a separate measure and a stated validity argument.

Report conclusion

Activity is not attention.

Method

Read the abstract and method summary

Web systems routinely report time-on-page, cursor activity, clicks, hovers, and similar signals as if they measured a person's attention. This report audits that practice on 2,909 crowdsourced web-search sessions with self-reported attention labels from the Attentive Cursor Dataset, of which 2,775 pass a minimal validity filter. We organize twelve common engagement proxies on a construct ladder from exposure to self-reported attention and measure how they relate to each other, to the participant's own attention report, and to context. The proxies disagree with one another: the median absolute pairwise rank correlation is 0.22, and raw dwell correlates at rho = -0.84 with the fraction of dwell that is active, so 81% of sessions land in discordant quadrants under a median split. Using the dataset's own focus, blur, and gap structure, we find that a median 51% of recorded dwell is idle or blurred, and 66% of sessions have at least 30% of dwell unattributable to observed, focused activity. Every proxy's association with self-reported attention is weak (absolute rho at most 0.063); a regularized multivariate model reaches AUC 0.558. Relationships shift across ad categories, including sign flips, and deterministic counterfactual transformations inflate headline metrics up to 59-fold without adding any evidence of attention. We distill the findings into an Attention-Claims Checklist for engagement reporting.

  • A construct ladder separating exposure, recorded presence, activity, continuity, orientation, and self-reported attention, with comprehension and intent explicitly off the ladder.
  • A quantitative disagreement audit among twelve engagement proxies computed on the same 2,775 sessions.
  • A direct measurement of raw-versus-active time inflation using the dataset's own focus and blur events.
  • A criterion-validity table relating each proxy to self-reported attention and to ad clicks, plus a participant-disjoint multivariate model reported with its modest performance.
  • Deterministic counterfactual stress tests quantifying metric gameability, and an Attention-Claims Checklist for anyone reporting engagement numbers.

Evidence and sources

  • The Attentive Cursor Dataset

    2,909 crowdsourced sessions of a search task on a results page carrying one ad; full client-side event logs; self-reported attention (1 to 5) and ad clicks. 2,775 sessions pass the validity filter.

    Label source and event record for the audit.

Limitations

  • The criterion is a single retrospective self-report item per participant: ordinal, coarse, and about noticing one ad, not attention in any richer sense.
  • One observation per participant: within-person questions are unanswerable here, and all context contrasts are between participants.
  • One desktop search task with a computer mouse; feeds, mobile, video, and multi-page journeys are out of scope.
  • The counterfactual results are statements about metric arithmetic under specified transformations, not estimates of how often such inflation occurs in the wild.

Not claimed

  • No log signal here measures comprehension, memory, interest, or intent, and no analysis supports claims about them.
  • Self-reported attention is an imperfect ordinal criterion, not ground truth about cognition.
  • All analyses are observational and correlational; nothing identifies causal effects.
  • The counterfactual stress tests describe no human behavior.

Reproducibility

All numbers are generated by deterministic analysis scripts (seed 42). Figures are produced from the same results file.

Citation

GrayPass (2026). Attention Beyond Activity: The Gap Between Action and Attention. GPL-TR-2026-03. https://www.graypass.org/research/attention-beyond-activity

Continue the work

Read the complete paper, then follow the next question.

The PDF contains the complete method, secondary results, limitations, references, and publication record.

Read GPL-TR-2026-03Next: The Shape of Identity