Date: 2026-04-01 Dataset: AdSERP (Latifzadeh, Gwizdka & Leiva, SIGIR 2025) Duration: Single session, ~2 hours from discovery to three notebooks
Frozen as of 2026-04-01 end of day. This document records the first session as it happened, including wrong turns and naive assumptions. Updates go in findings.md. See "What We Got Wrong" at the end for corrections applied in v1+.
The session started with a paper drop — Andy spotted "A Versatile Dataset of Mouse and Eye Movements on Search Engine Results Pages" (SIGIR '25). Open access. 2,776 transactional queries on Google SERPs, 47 participants, simultaneous Gazepoint GP3 HD eye tracking (150 Hz) + evtrack mouse capture. The dataset includes SERP HTML snapshots, fixation data, mouse trajectories, scroll events, pupil dilation, and programmatic ad bounding boxes.
The immediate connection: this dataset sits at the intersection of Scrutinizer (gaze replay, foveated rendering) and ClickSense (click commitment dynamics). Gaze replay through Scrutinizer's pipeline on real SERP scanpaths would show what peripheral vision actually delivers at each fixation.
Andy's first observation: academic work reporting mouse-gaze divergence (Huang, White & Buscher 2012; the AdSERP paper itself reports 372px mean distance) has never conditioned on p(click). The higher the click probability, the closer the mouse should be to the eye — because the mouse must converge on the gaze target to execute the click. Averaging across entire browsing sessions (including scanning/idle periods) inflates the reported divergence.
This is a simple idea. Huang, White & Buscher (2012) showed gaze-cursor alignment tightens near clicks, but the continuous distance-as-a-function-of-time-to-click curve — and the scroll correction that changes its shape — hadn't been done.
Setup: Cloned the AdSERP repo, downloaded behavioral data from Zenodo (~15MB for fixation + mouse + metadata). Built convergence.py — for each of 2,776 trials, compute mouse-gaze Euclidean distance at each fixation, tagged by time-remaining-to-click.
First run (broken): 12,556 pairs, 2,269 trials skipped. The fixation data had float values where we expected ints. Fixed parsing, reran.
Second run: 128,887 fixation-mouse pairs across 2,762 trials. The hypothesis holds:
| Time to click | Mean distance (px) |
|---|---|
| 20-60s | 330 |
| 10-20s | 334 |
| 5-10s | 314 |
| 3-5s | 282 |
| 2-3s | 240 |
| 1-2s | 172 |
Monotonic decrease. 48% reduction from scanning baseline to acquisition. The reported 372px aggregate sits above the scanning baseline because it includes all the zero-intent noise.
Three phases emerged:
- Scanning (>5s): mouse parked, eye foraging. ~330px divergence.
- Evaluation (2-5s): eye on candidate target, mouse drifting toward it.
- Acquisition (0-2s): motor commitment, mouse converging. ~172px.
Andy identified the 1-2s window as the target acquisition phase (consistent with ClickSense trajectory data) and noted the 2-5s evaluation window preceding it as the interesting zone — the duration is variable and could be calibrated per session by result evaluation time.
The 0.5-1s bump: Distance spikes briefly at 0.5-1s before click before collapsing to click. Consistent with a verification saccade — the eye breaks from the target to check something peripherally before committing.
Per-participant variation: Acquisition onset ranges from 0.2s to 13.8s (mean=2.4s, SD=2.5s). Individual differences are large enough to warrant per-session calibration.
X-Y decomposition: X divergence dominates during scanning (mouse parked to the side, 245px vs 180px Y). The ratio inverts near click — Y briefly exceeds X at the verification window (ratio 1.12). SERPs are vertical layouts; the mouse approaches horizontally first, then the vertical component catches up for the final commit.
Click prediction: Logistic regression at 5s horizon gives AUC=0.638. Peak at 2s horizon (AUC=0.720). Raw distance is the dominant feature; convergence rate (slope) adds only +0.002 AUC. The distance signal is simple and powerful; the derivative doesn't help a linear model.
Checking event types in the mouse data revealed scroll events — window.scrollY offsets captured at ~60Hz. They're the most frequent event type in many trials. This changed everything.
Fixation coordinates are in page space (gaze Y goes to 1647 on a 1024px screen). Mouse coordinates are in screen space. The scroll offset is the missing translation between coordinate systems — but more importantly, scroll state is a behavioral signal.
The viewport chart (Andy flagged this as critical): the click target is in the viewport only ~50% of the time before 10s — essentially chance. The target-in-viewport rate only rises meaningfully in the last 10 seconds, hitting ~66% at 5s and ~75% at 2s.
Andy's reframe: Before 10s, "distance to click target" is measuring distance to something often not on screen. It's an abstract distance-to-goal metric, not a spatial-motor signal. The convergence story only becomes spatially meaningful once the viewport locks onto the target region.
Scroll features beat everything: Adding viewport state to the logistic regression jumped AUC from 0.631 to 0.704 (+0.074). Scroll features alone (0.687) beat distance+all-prior-features (0.638). Target-in-viewport is the key predictor. The convergence signal is downstream of the scroll-stop event.
Andy's reaction to seeing scroll regressions in the data: "we have regressions — that's huge, I'm unaware of much work in research quantifying those."
Eye movement regressions in reading are a massive literature. Page-level regressions — scrolling back up to re-examine previously viewed results — are barely studied.
Prevalence: 69.1% of trials contain at least one scroll regression. This is the norm, not the exception.
Magnitude: Mean 2.8 regressions per trial, mean regression distance 1,118px (~7 result slots). Users don't just glance back one slot — they traverse significant portions of the SERP.
Timing: Regressions cluster in the middle of the trial, not at the end. They're part of the evaluation process, not a last-minute correction.
Decision cost: Trials with regressions take 11.9 seconds longer (21.6s vs 9.7s). Regression count correlates with decision time at r=0.660.
Per-participant: Regression rate varies from 11% to 98% across participants (mean 66.8%, SD 20.6%).
Sparkline visualization: 20 highest-regression trials show a characteristic "mountain" pattern — scroll down, peak, scroll back up, settle, sometimes repeat. The scroll position trace is a decision narrative.
Andy's hypothesis: the well-documented finding that users evaluate results faster as they scroll down is attributed to decreasing effort or attention fatigue. The alternative: it's cumulative lexical priming. After reading results 1-3 for a query, the user has accumulated enough shared vocabulary that subsequent results can be evaluated faster — not because they care less, but because the context model is richer.
Extracted text from 2,772 SERP HTML files using BeautifulSoup. For each result, computed token set and cumulative overlap with all preceding results.
The priming curve is strong:
- Position 1: 38% of tokens already seen in position 0
- Position 5: 56% overlap
- Position 9: 62% overlap
- Novel tokens per result drop from 28 (position 0) to ~10 (position 9)
By the time a user reaches result 9, nearly two-thirds of its vocabulary is already primed. The information environment becomes increasingly redundant.
SERP-level homogeneity doesn't predict regressions (r=-0.015). This null result is informative: regressions aren't driven by how homogeneous the overall page is. They're likely triggered by local novelty events — a single result that breaks the pattern, introducing vocabulary that doesn't fit the accumulated context. That forces re-evaluation of prior results in light of new information. This per-result novelty → regression trigger analysis is the next step.
Andy also noted that the priming hypothesis connects to his eBay research: regressing time-to-first-click against image size showed bigger images add ~500ms evaluation time. Viewport slot economics — high-information elements cost more to evaluate but the cost is modulated by what context has already been established.
- p(click) conditioning — mouse-gaze distance is a function of decision state, not a fixed property
- Two-regime distance metric — before 10s it's abstract distance-to-goal; after 10s it's spatial distance-to-visible-target
- Scroll-stop as the real predictor — viewport state (AUC=0.704) beats gaze-mouse distance (0.631) for click prediction
- Scroll regressions are the norm — 69.1% of trials, barely studied in the literature
- Lexical priming curve — cumulative overlap reaches 62% by position 9 (a content-structural fact; the behavioral claim that this explains acceleration did not survive within-position controls — see "What We Got Wrong v2–v4")
- Individual calibration — 2.5s SD in acquisition onset; 20.6% SD in regression rate
The authors' contribution is the dataset itself — a generous, well-structured public resource. These are directions we pursued, enabled by the richness of what they provided:
- Conditioning on click intent and decision stage
- Temporal dynamics (mouse lag/lead vs gaze)
- Scroll regression characterization
- SERP content analysis using the provided HTML snapshots
- Pre-attentive processing and peripheral vision questions
- Per-participant variability analysis
- Scrutinizer: Gaze replay on SERP scanpaths. Mouse position replay added to TODO. Foveated rendering would show what peripheral vision delivers at each fixation — which elements are detectable pre-attentively.
- ClickSense: Scroll regressions as macro-scale hesitation; click hold duration as micro-scale. If they correlate, that's a multi-timescale confidence measure.
- evtrack: Same author (Leiva) behind both the capture library and the dataset. ClickSense is the next-generation capture — narrower but deeper.
- Per-result novelty → regression trigger: Does a specific high-novelty result (low cumulative overlap) predict a regression at that scroll position?
- Pupil dilation × regressions: The dataset includes pupil data. Pupil dilates on cognitive load/surprise — does it spike during regressions?
- Fixation-to-result mapping: Use ad bounding boxes + result positions to assign each fixation to a specific SERP result, enabling per-result evaluation time analysis.
- Scrutinizer replay: Render foveated view at each fixation on the actual SERP screenshots. What information is available peripherally at each moment?
- Interactive visualizations: The matplotlib plots are a first pass. The sparklines and convergence curves deserve D3/interactive treatment.
- Coordinate correction: Properly reconcile page-space fixations with screen-space mouse coordinates using scroll offset. Current analysis uses uncorrected screen-space distance — the scroll correction may sharpen the convergence signal in the 5-10s window.
attentional-foraging/
├── convergence_analysis.ipynb — Notebook 1: convergence + prediction (12 plots)
├── scroll_regressions.ipynb — Notebook 2: regression characterization (3 plot sets)
├── serp_priming.ipynb — Notebook 3: lexical overlap + regression link (2 plot sets)
├── convergence.py — Standalone aggregate analysis
├── phases.py — Per-trial phase detection (not yet integrated)
├── docs/
│ ├── journey.md — This document
│ └── adserp-key-claims.md — Theoretical gaps in the original paper
├── html/ — Exported notebooks for sharing
│ ├── convergence_analysis.html
│ ├── scroll_regressions.html
│ └── serp_priming.html
└── AdSERP/ — Cloned repo + Zenodo data
└── data/
├── fixation-data/ — 2,776 CSVs (page-space coordinates)
├── mouse-movement-data/ — 2,776 CSVs (screen-space + scroll + click)
├── trial-metadata/ — 2,776 XMLs (viewport dimensions, query)
├── serps/ — 2,776 self-contained HTML files
└── participants.csv — 47 participants, demographics
Coordinate mismatch (v0 → v1): The v0 convergence plots used uncorrected screen-space coordinates for both gaze and mouse. Fixation Y is in page-space (increases past screen height as user scrolls); mouse Y is in screen-space. The scroll offset needed to reconcile them was available from the start but not applied until v1.
The corrected plot tells a different story: distance starts low (both gaze and mouse near top of page), increases as the user scrolls (gaze follows content down the page, mouse stays in screen space), then converges sharply in the last few seconds before click. The 372px aggregate sits in the middle of the curve rather than above the baseline.
The relative findings (two-regime model, convergence timing, viewport features beating distance) are robust — they depend on trends, not absolute pixel values. But the absolute numbers in the v0 table above are wrong.
Click prediction AUC (v0 → v1): Distance-only baseline dropped from 0.631 to 0.548 with corrected coordinates. Scroll features (0.687) and the full model (0.704) held. The distance signal was inflated by the coordinate mismatch — scroll features were doing most of the work all along.
During a follow-up session, these findings were connected to the Attentional-Foraging Equilibrium — Andy's framework synthesizing Rational Inattention (Sims) with Information Foraging Theory (Pirolli & Card). The AFE equation ρ = V / (τ + T_s + σ²) maps directly onto the findings: lexical priming reduces τ, scroll regressions are T_s costs, the convergence curve traces foraging-to-exploitation transition, and per-participant variance reflects bandwidth λ. The forced-choice purchase task turns out to be a serendipitous paradigm — it makes the patch-leaving decision observable. See the updated findings.md for the full mapping.
The priming headline collapsed under scrutiny. The v1 story was: "lexical priming predicts faster evaluation (partial r = -0.054)." Each subsequent revision peeled away a layer of confound:
- v2: Regression-stratified analysis. The effect was concentrated in regression trials (r = -0.033), null in first-pass (r = -0.002). Reframed as "priming facilitates re-evaluation, not first-pass scanning." This felt like a sharper, more honest finding.
- v3: Within-position controls. Testing high-overlap vs low-overlap at the same rank — null across every metric (TFT, TFC, mean fixation duration, viewport time). The aggregate correlation was driven by the position-overlap confound: both decline monotonically with rank, so any position-correlated attention decline looks like a priming effect.
- v4: Forward-only shape test. Isolating forward-scanning periods (excluding regressions), gaze dwell ratio increases with position (Spearman ρ = +0.73) — the opposite of the priming prediction. The "priming effect" was entirely an artifact of regressions (lower dwell on revisit = memory, not priming) plus position confounding.
The metric naming was sloppy. We called it "eval rate," then "attention density," then "gaze dwell fraction." Each name was imprecise about what was being measured: the ratio of two durations (fixation duration / visible duration). "Gaze dwell ratio" was the final name. Getting the name wrong made it harder to think clearly about what the metric could and couldn't tell us.
The viewport time computation was broken. compute_viewport_time only counted intervals between scroll events. Before the first scroll, the viewport is static at y=0 — position 0 is fully visible — but that entire period was dropped. Some trials had 27 seconds of fixation on 183ms of computed viewport time, producing dwell ratios of 73x. 44% of position-0 observations had ratios >1.0. We reported means built on these numbers.
We treated per-fixation duration as a priming metric. The ~220ms flat-across-positions finding was presented as "reframing the priming question" — implying fixation duration was a reasonable thing to test. It wasn't. Fixation duration is a low-level oculomotor parameter. No one in reading research would predict that result-level vocabulary overlap changes individual fixation durations. The grain size is wrong by an order of magnitude. This is a cognitive psychologist's error — confusing the level of analysis. The valid priming metrics were always fixation count and p(fixate), both of which are null within-position.
We tested p(fixate | visible) and found it structurally uninformative. Forward-only p(fixate) is ~99.8% at every position. During first-pass scanning, users fixate everything they see. There is no skip decision for overlap to predict. The 12.5% skip rate lives entirely in regressions and late-trial edge cases. This was the last plausible refuge for a bag-of-words priming signal, and it's empty.
The pattern across v1–v4: Each version reported a "finding" that looked significant because we hadn't controlled for the right thing yet. The aggregate r survived because position and overlap are confounded. The regression split survived because regressions conflate familiarity with revisitation. The within-position test was the real control, and it's null. The forward-only shape test confirmed the null by showing the curve goes the wrong way. We should have started with within-position controls and forward-only isolation. Instead we reported the exciting aggregate and spent four versions discovering it was confounded.
During gazeplot overlay alignment work, discovered a subtlety in the recording setup:
Window exceeds screen. Trial metadata reports window: 1422x1137 on screen: 1280x1024. The browser window is 142px wider and 113px taller than the physical display. On Windows, maximized browser windows extend beyond the screen edges (the frame/shadow is off-screen). The visible content area is smaller than windowWidth.
FPOGX/FPOGY are page-space, confirmed by Y values exceeding screenHeight (max pageY=1833 in p037-b2-t5 on a 1024px screen). The Gazepoint GP3 HD eye tracker reports gaze in screen pixels (0–1280 × 0–1024). The conversion to page-space requires adding the scroll offset — either the Gazepoint SDK was configured to do this, or it was applied in post-processing. The fixation CSV column names (FPOGX/FPOGY) follow Gazepoint naming conventions but the values are definitively page-space.
The coordinate mapping for gazeplot overlays:
- FPOGX is page-space X at the window-width layout (1422px). Range observed: 190–796.
- FPOGY (pageY) is page-space Y at the window-width layout. Range: 17–1833.
- For SVG overlay on a 1422px-wide SERP render: X needs screen→window scaling (
× windowW/screenW), Y maps directly (pageY is already in the 1422px layout's coordinate system). - ~30px vertical drift at the bottom of long pages (2359px). Likely cross-browser rendering difference between the original Chrome 110 (Windows) and Playwright's Chromium (macOS). Within tolerance for the visualization but worth documenting.
Resolved (same session). The AdSERP README states explicitly: FPOGX/FPOGY are "relative to the top-left corner of the screenshot in pixels." The screenshot is at screenWidth (1280px). This means:
- FPOGX/FPOGY map directly to a 1280px-wide full-page screenshot. No scaling needed.
- We had been rendering SERPs at
windowWidth(1422px) and applying screen→window coordinate transforms. This was wrong — the 30px drift at the bottom of the page was reflow from rendering at the wrong width, not a cross-browser difference. - Fix: Render SERPs at 1280px (screen width). Map FPOGX/FPOGY directly. No coordinate transforms.
- The
window: 1422x1137metadata is the browser window outer size (extends beyond the 1280×1024 physical screen on Windows). It's irrelevant for fixation coordinate mapping.
Git repo made public 8:38am PT, 2026-04-01. ~3.5 hours from dataset discovery to public release.
v0 frozen 2026-04-01. Corrections appended as they were discovered. The dataset is rich enough for a full paper.