review methods and limits (authored by agents unless marked 🧑)
scope
- extensive targeted literature review across four user-facing topics
- 19 search-quality papers plus Google policy, 21 phishing/scam studies, 15 full-text ad studies, 14 full-text dependency studies
- counts are track entries, not deduplicated papers
- overlapping studies connect topics
- abstract-only and metadata-only leads are marked
- recommendations are proposed experiments
- no measurements of today’s search quality or spam prevalence were performed
- no claim of exhaustive bibliography or established novelty
human context read first
- web user-facing
- research goals
- removal services
- DeGenTWeb ad extraction, site classification, non-article filtering, and Google Trends notes
- the initial linked files contained titles only
- final group review found substantive project classification and filtering notes
- task focus, quoted from the human’s research index
- “how to find knowledge among spam on the web”
source discovery
- follow references from the human’s notes
- inspect publisher and author pages, venue programs, arXiv, and available paper-collection files
- retrieve primary PDFs and extract text
- record exact source locations for quotations
- follow relevant citations and related studies
- search access limits
- broad search services and some publisher full texts were unavailable
- direct primary pages, venue programs, and collection PDFs supported the review
- consequence
- exact-term arXiv searches later succeeded
- recent addenda review closer work on copied evidence, citations, ads in answers, and fine-grained blocking
- reliable direct evidence is stronger than the completeness of recent-literature coverage
- recent USENIX programs and exact-term arXiv results were inspected where useful
- comparable complete scans across all venues were not possible
paper notes
- identify question, population, method, result, and limitation
- keep short original quotes with exact author and source pointers
- distinguish source findings from our interpretation
- keep historical percentages within their original denominators
- flag source conflicts rather than silently reconcile them
- Free Waters has coverage/date ambiguities
- Africa dependency PDF is an anonymous draft
- one ad paper is abstract-only because its retrieved PDF is truncated
- rendering-dependency poster is metadata-only
archive
- PDFs saved locally under
/hdd1/sichanghe/paper_collection - collection repository ignores PDFs
- source records and extracted text committed and pushed
- 54 PDFs retained locally in the first archive pass
- 12 additional PDFs preserved for recent-work addenda
- new source records distinguish URL and retrieval date
- paper bytes checked before saving as PDFs
- track source pages point to the corresponding collection folders
- source records and PDFs are separate from review prose
- permits later full-paper reading without relying on these summaries
review
- independent context-free reviewer checked the three completed topic tracks and dependency literature
- selected statistical and quotation claims checked against primary PDF text
- corrections applied
- DynaPhish change expressed in percentage points
- PhishDecloaker result retains the combined study denominator
- copied-evidence experiment keeps evidence fixed and controls repetition separately from source identity
- reading counts and archive provenance reconciled
- review limits
- recent addenda independently reviewed against preserved PDF text
- corrections distinguish attack conditions and audit denominators
- no independent check of every paper sentence
- no exhaustive verification that proposals are new
ChatGPT consultation
- the saved GPT-6.1 Sol consultation used Extra High on 8 Oct 2026
- its finished answer was recovered from ChatGPT’s conversation record
- the page remained stale after completion
- assessment and resulting changes
- no ChatGPT opinion is treated as literature evidence
research measurement rules
- separate website, visit, ad, campaign, query, claim, and user denominators
- sample by the outcome being claimed
- popular websites do not stand for all user tasks
- controlled tasks do not establish population prevalence
- archive configurations and failures
- independent annotation with disagreements retained
- separate current facts, outdated facts, unsupported claims, and uncertain origin
- separate commercial incentive from factual error
- compare simple baselines before complex classifiers
- preserve natural browsing runs separately from forced execution
- controlled experiments support mechanism claims
- web-wide prevalence requires a justified sample
follow-up reading priorities
- final Web Dependency Analyzer text and artifact
- source dependence, citation laundering, and retrieval poisoning
- current browser storage and breakage studies
- broader 2026 search-quality and ad-measurement venue coverage
- separate people-search removal review
- phishing removal results cannot establish people-search removal effectiveness
Last edited: