A Cue's Hidden Price: Auditing Seller Influence on AI Shoppers
The paper introduces price-equivalent auditing to measure how marketing cues steer AI shoppers in dollars rather than saturated choice rates.
本站有中文版。
切换到中文We watch arXiv for papers about how AI engines choose and cite sources. Each write-up says why the paper matters for getting cited, quotes the abstract and key passages as written, explains the terms, lists what the authors say it means, and ends with what to do and what the paper did not test. Checked by a person before it goes up. Our own papers are on the research page.
The paper introduces price-equivalent auditing to measure how marketing cues steer AI shoppers in dollars rather than saturated choice rates.
The authors propose a three-part taxonomy and benchmarking protocol to standardize how generative engine optimization strategies are defined, constrained, and evaluated across AI systems.
ChatGPT personalizes sources by declared city while Gemini localizes by network country, significantly altering cited domains and business recommendations.
The authors evaluate generative engine optimization in elections and show that page structure and argumentation boost target exposure across voter personas.
An observational study of ChatGPT Web shows that retrieval differs across model settings and that combined external search proxies recover only about one quarter of cited URLs.
Across 158,099 generative search answers, retrieval explains far more citation variation than page features, while common generative engine optimization tactics fail within the same site.
A field experiment with 37,561 Washington Post readers shows that generative AI search increases shared topic consumption while spreading attention to less popular topics.
The paper shows that early citations in conversational search compound across turns, causing single-turn Generative Engine Optimization evaluations to misjudge cumulative visibility.
AI agents rely on live web fetches rather than memory, making site readability far more decisive than third-party citations.
Single AI answers identify their generating system through surface formatting, but aggregated brand recommendation profiles fail to transfer across domains.
This audit reveals that Baidu and Google AI overviews produce semantically similar answers while citing substantially different and concentrated sets of host domains.
An audit shows AI search engines rarely cite the same URLs for identical prompts, limiting the reach of standalone page scores.
An audit of 1,536 chatbot responses reveals that AI recommendations vary across repeated queries and that consumer web interfaces cite different sources than developer APIs.