AllSERP — Resource Paper Repository

AllSERP: Exhaustive Per-Element Enrichment of the Versatile AdSERP Dataset Andy Edmonds. arXiv:2605.04949.

Live on arXiv: v3 (2026-07-23). v4 is drafted, rebuilt, and bundled (2026-09-07), and clear to upload. The change-note for each version is in CHANGELOG.md.

---

What the paper is

A resource / dataset-enrichment paper. AdSERP ships 2,776 search trials with gaze, cursor, scroll, pupil and click telemetry on real Google result pages, but its bounding boxes cover only advertisements. AllSERP adds a typed area-of-interest (AOI) layer for everything else: screenshot-anchored boxes for organic results, People-Also-Ask, image packs, local packs and other widgets, labelled from the captured HTML. The paper describes the pipeline, validates it, reports a per-element behavioural inventory, and documents how coverage grew across and within element types.

It is not a model paper. Downstream model work lives in sibling tracks and appears only as "what's enabled" pointers.

Headline claims (v4, on AllSERP v1.1.1)

1. Ad partition validated against the shipped ad rectangles: 0 disagreements across 38,250 classifications. A DOM-based harness measures box fidelity on the full corpus.
2. Three released flavours trade box tightness for coverage: typed (tight boxes), typed_gapfill (inter-result Y gaps filled by midpoint split), and typed_gapfill_cellsplit (per-card cells inside the top-ads carousel).
3. 95.7 % of final clicks land in a typed main-axis AOI under typed_gapfill once evtrack cursor coordinates are converted into screenshot space (91.5 % on the shipped document-space filter). The shipped is_main_axis_click helper is still document-space; the paper discloses this as a queued substrate revision.
4. Calibration-bias hypothesis tested and refuted: opposite-direction click vs fixation bias confirms the data is screenshot-aligned.
5. Per-element inventory across 8 main-axis element types (organic, dd_top, native_ad, paa, image_pack, top_places, unknown_widget, other_widget): click share, fixation coverage, regression rate, above-fold incidence. Knowledge panels are right-rail (position −1) and no longer a main-axis row.
6. Coverage section (new in v4): how coverage grew across element types (organic boxes → HTML typing → gap-fill → geometric label verification in v1.1.0) and within them (the carousel cell layer). The v3 within-carousel rank claim (ρ = −1.0) is withdrawn and restated as directional; the released cells come from a May 2026 snapshot and the paper says so.

Substrate identity. Every number is pinned to AllSERP enrichment v1.1.1: 2,764 analysable trials, a 12-trial alignment-exclusion list, typed-map content hash 2cb789eb8febd234. A stale export contains ~746 main-column knowledge-panel rows and 84 local packs; v1.1.1 contains 0 and 340.

---

Layout

allserp-paper/
  paper.tex               CANONICAL source (ACM acmart). Build: ./build.sh --acmart
  paper-acmart.pdf        tracked build output, for early sharing
  paper.md                FROZEN legacy markdown draft (pre-2026-05-06); not current
  build.sh                acmart build (xelatex + bibtex); legacy pandoc paths retained
  make-arxiv-bundle.sh    rebuilds, then stages paper.tex/.bbl/bib/figs -> allserp-arxiv.tar.gz
  bib/allserp.bib         bibliography
  figs/                   figures + CAPTIONS.md + render_*.py / render_replay.js producers
  CHANGELOG.md            per-arXiv-version change-notes and decision history
  TODO.md                 v4 worklist and the pre-upload landing checklist
  CLAUDE.md               editing conventions (citation discipline, voice, commit types)

Not tracked: texmf/ and usertexmf/ (the local acmart install that build.sh expects via TEXMFHOME), .arxiv-staging/, the tarball, working PDFs, and notes/ (private coordination). A fresh clone needs acmart installed locally before ./build.sh --acmart will run.

Building

./build.sh --acmart          # paper.tex -> paper-acmart.pdf
./make-arxiv-bundle.sh       # rebuild + stage -> allserp-arxiv.tar.gz

Figures regenerate from figs/render*.py and figs/render_replay.js; they read the attentional-foraging outputs listed below. figs/CAPTIONS.md holds the caption text.

---

Sources of truth

The paper does not re-derive numbers. Every quantitative claim traces to a file in one of two repositories:

• attentional-foraging — pipeline and producers. Substrate branch release/allserp-v1.1.0, tag allserp-v1.1.1 (GitHub release with the four corpus CSVs and the exclusion list attached).
  • scripts/allserp_descriptives.py --flavor typed_gapfill --space screenshot — Table 1
  • scripts/audit_*.py — cite-ready audit producers (audit_cascade_contamination.py, audit_dd_right.py --space screenshot, and the ad-rectangle check)
  • scripts/output/adserp_aois_by_trial_id_{typed,typed_gapfill,organic_hybrid,typed_gapfill_cellsplit}.csv and data/aoi-typed/alignment-exclusions.json — the released exports
  • docs/methodology/organic-result-aoi-extraction.md — pipeline spec
  • docs/methodology/attribution-cascade-synthesis.md — flavour history
  • docs/methodology/dd-top-cellsplit.md — cell-split tiers, dd_right-as-covariate
  • docs/methodology/carousel-full-corpus-validation.md — DOM-derived carousel candidate (1,570/1,575 admitted, 7,265 cards)
  • docs/allserp-v1.1.0-migration.md, docs/releases/allserp-v1.1.1.md — substrate notes
  • docs/notebook-key-claims.md — Key Claim IDs cited in the text
• approach-retreat — the replay viewer (148 curated trials, labelled log-scaled LF/HF tracks as of the v4 figure).

The underlying corpus is the AdSERP Zenodo volume (zenodo.org/records/15236546, CC-BY-4.0). AllSERP does not redistribute it.

---

Editing

Conventions for edits (two-pass citation discipline, voice, sibling-track guard, conventional-commit types) are in CLAUDE.md. Short version: no author names or venue tokens in citation position until the source has been verified against its abstract, and no number without a [<flavor>, <source>] provenance.

History

date	event
2026-05-05	repository scaffolded from the bbox-y-coverage-fix cascade
2026-05-06	arXiv v1 submitted
2026-05-19	arXiv v2: ARS-audit / reviewer-pass cleanup
2026-06-23	paper.md frozen; paper.tex canonical
2026-07-23	arXiv v3: dd_top cell-split enrichment, within-carousel ordering
2026-09-07	v4 drafted and bundled: restructure, re-derivation on v1.1.1, Coverage section; upload pending

Citation

@misc{edmonds2026allserp,
  title  = {AllSERP: Exhaustive Per-Element Enrichment of the Versatile AdSERP Dataset},
  author = {Edmonds, Andy},
  year   = {2026},
  eprint = {2605.04949},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/2605.04949}
}