Seller message framing tracked purchases in live persuasion games
Source check: Source checked. Not independent scientific confirmation.
Finding: A retrospective analysis of 9,337 classifiable messages from one AI seller in 725 GLEE persuasion games found that buyers purchased after 65.8% of loyalty-framed messages and 30.9% of care-framed messages. The association persisted when the study adjusted for product quality and game round.
Technology: Online Thompson sampling bandit; language-model message generation
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- An online Thompson sampling bandit fitted framing success probabilities during play and supplied the framing for the language model that wrote the analyzed seller messages.
- Test and result
- Among 9,337 lexicon-classifiable outgoing messages in 725 live persuasion competition games, buyer response was 65.8% for loyalty-framed versus 30.9% for care-framed messages. Game-clustered, quality- and round-adjusted odds ratio was 3.98 (95% CI 3.31–4.79), or 2.18 (1.78–2.67) for pitches only.
- Limitation / still unknown
- The agent chose framing adaptively and labels came from a lexicon. Opponents included humans and AI agents, and the aggregate does not identify a human-only effect. Repeat opponents could not be accounted for, and framing was not randomized. About half of care-labelled messages were honest recommendations not to buy, compared with 6–17% for other frames, so the raw purchase-rate contrast also reflects message intent. The pitches-only restriction reduces this difference but is itself a lexical approximation.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/html/2610.37397v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review date: 21 Nov 2026. Review scope: Versioned primary arXiv HTML front matter, §§3.3, 4 and 6, Appendix D Tables 7, 8 and 12.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.