Steven Hé (Sīchàng)

recent work on advertising inside answers and children’s websites (authored by agents unless marked 🧑)

what changes the proposed research

• inference: inserting and evaluating advertisements inside AI answers already has substantial recent work
  • the strongest remaining proposal here is measuring loss of an existing sponsorship label across representations
  • building another ad insertion method would overlap these papers
• evidence status
  • full PDFs, methods, evaluation, and limitations inspected on 2026-10-07 UTC
  • the two 2026 papers are arXiv preprints
  • the children’s-websites study appeared at IEEE S&P 2024

Evaluating and Pricing Advertisements in AI-Generated Responses

• John L. Turner-Smith, Zimeng Huang, Yuhan Fu, Yihang Zhang, Tonghan Wang
  • inspected version: arXiv v1, July 2026
• question: can simulated reader preferences provide a score usable for choosing and pricing advertisements?
• method
  • train an evaluator using 58,999 advertisement-containing responses from NaiAD
  • share a learned representation across relevance, coherence, advertising effectiveness, and click intent
  • create click-intent labels from model-assessed features and mathematically sampled personality traits
  • test controlled changes to content and relevance
  • derive auction pricing assuming an allocation that rises with bids
• reported evaluation, §4
  • 5,837 held-out items
  • relevance swaps lower the predicted intent for 11 of 14 non-tied comparisons
    • the reported 79% rate comes from this small test
  • five annotators each judge 100 response pairs
    • 92% agreement with majority choice
    • 86% mean agreement with individual annotators
  • tests on 103 fictional products avoid memorizing existing brand names
• decisive limitation, original words, §6
  • “an ordinal surrogate for engagement, not a calibrated probability”
  • context: the evaluator’s 1–5 score does not measure observed click probability
• inference: consistency with assumed behavioral directions cannot validate actual persuasion
  • some personality effects enter the labels by construction
  • human comparisons judge likely clicks rather than record clicks
  • pricing results illustrate a mechanism under its assumptions
    • they do not establish realized revenue
• relevance to our proposal
  • compare evaluator scores with whether an answer preserves commercial attribution
  • keep disclosure retention, factual accuracy, and willingness to click as separate outcomes
  • avoid adopting synthetic click intent as ground truth for user welfare

Ad Insertion in LLM-Generated Responses

• Shengwei Xu, Zhaohua Chen, Xiaotie Deng, Zhiyi Huang, Grant Schoenebeck
  • inspected version: arXiv v2, August 2026
• method
  • generate ordinary answer text before inserting screened advertisements
  • advertisers bid on broad semantic categories called genres
  • score how well each genre fits each insertion point
  • use a VCG auction
    • winning advertisers pay for the value their allocation removes from other bidders
• original words, abstract
  • “decouple ad insertion from response generation”
  • context: this permits separate screening and explicit disclosure
• reported evaluation
  • 48 exploratory survey participants in China
  • 36 complete the coherence study
    • seven prompt contexts and ten ad genres
    • GPT-5 coherence judgments correlate with mean human ratings at Spearman ρ ≈ 0.66
  • synthetic auction with 100,000 advertisers and 100 slots clears in about 1.25 seconds
• limits, §§4, 6, 8
  • coherence signals still need calibration against observed interaction data
  • prototype score normalization does not establish the calibrated probabilities assumed by the guarantees
  • highly educated participants and seven contexts restrict generalization
  • survey concerns and coherence ratings do not measure clicks, misinformation, or retained sponsorship knowledge
• inference: useful baseline for separating ordinary text from sponsored inserts
  • compare a disclosure-preserving format against promotion blended into answer text
  • measure whether downstream summarization preserves that separation
  • this study’s explicit insertion disclosure does not settle recognition after later transformations

Targeted and Troublesome: Tracking and Advertising on Children’s Websites

• Zahra Moti, Asuman Senol, Hamid Bostani, Frederik Zuiderveen Borgesius, Veelasha Moonsamy, Arunesh Mathur, Gunes Acar
  • inspected version: arXiv v2, December 2023
  • IEEE S&P 2024
• method
  • fine-tune a multilingual classifier on page titles and descriptions
  • search 2.28 million Common Crawl pages and manually verify selected candidates
  • final list: 2,004 child-directed websites across 48 languages
  • seven April 2023 crawls
    • five desktop locations and two mobile locations
  • EasyList selectors locate ad elements and their descendants
  • scrape Google and Criteo disclosure pages to classify the disclosed targeting setting
  • recognize text in ad images and rank it by semantic similarity to topics
    • inspect the top 100 distinct passages per selected topic
    • use an image classifier to identify sexually suggestive material
• reported results, §5
  • about 90% of listed sites contain tracker domains
  • targeted ads occur on roughly 27% of sites
  • 73% of ads with readable disclosures indicate targeting
    • this denominator excludes ads without supported disclosure information
• original words, §6.3
  • “only used ad disclosure pages from two providers”
  • context: targeting measurements depend on Google and Criteo’s disclosure coverage and accuracy
• extraction validation, §4
  • manual check of 105 extracted items finds 85% valid ads
    • 7.5% ordinary content and 7.5% empty slots
  • review of 50 pages without detected ads finds no missed ads
    • this small check does not establish universal recall
• limits
  • classifier favors precision and may favor English or descriptively titled sites
  • high-score selection and manual validation do not create a representative sample of every children’s site
  • some mixed-audience sites remain
  • consent-dialog behavior can confound geographic comparisons
  • disclosure indicates a provider’s stated reason
    • it does not reveal every internal decision or a real child’s observed exposure
  • top semantic matches demonstrate concerning examples
    • they do not establish complete prevalence of every harmful category
  • historical legal discussion should not be treated as current legal advice
• contribution relevant here
  • supplies a tested combination of slot extraction, disclosure collection, text recognition, and human review
  • supports a follow-up on whether accessibility tools and answering systems retain advertising disclosures on educational pages

further leads found in the 2026 papers

• these were identified in references, not fully reviewed here
  • NaiAD: Initiate Data-Driven Research for LLM Advertising, arXiv:2605.09918
  • GEM-Bench, arXiv:2509.14221
  • PILA, arXiv:2607.25590
  • LLM-Auction, arXiv:2512.10551
  • Ads that Talk Back, Tang et al., IMWUT 2025
• recommendation: check these before asserting that an answer-advertising benchmark is new

Last edited: 2026-10-06