Whose hotel does the AI recommend? An algorithm audit of reputation signals in LLM-assisted hotel selection

Large language model assistants that recommend hotels weigh guest rating and price above all other signals, according to a new algorithm audit. The study also found that a hotel’s position in a list — a factor unrelated to quality — causally shifts recommendations by an amount equivalent to roughly $12 per night. The pre-specified audit, posted to the arXiv preprint server, used a randomized choice-based conjoint design across twelve open-weight and proprietary models [1][2]. Researchers independently varied guest rating, review volume and recency, management response, chain affiliation, price, eco-certification, and list position for sets of five hotels, then estimated the average marginal component effect of each signal on the probability of being recommended [2]. A top guest rating raised selection by 31.6 percentage points, while a high price lowered it by 30.0 percentage points [2]. The dominance of valence and price mirrors patterns observed in human decision-making, the authors note, but the models over-weighted eco-certification and ignored management response [2]. List position — a content-free artifact — exerted a causal effect on recommendations worth about $12 per night [2]. The stated reasons the models gave for their choices tracked the revealed weights only imperfectly [2]. The paper, authored by Mirza Samad Ahmed Baig and submitted on 15 June 2026, frames the results as grounding for generative engine optimization and the accountability of AI infomediaries [1][2]. Large language models are a class of machine learning systems trained on vast text corpora for tasks such as language generation, and they have been deployed across sectors including e-commerce and decision-support [3][9]. The audit’s findings arrive as travelers increasingly turn to these assistants for lodging advice, making the systems de facto gatekeepers of property visibility [2]. The study did not examine whether the observed patterns vary across geographic markets or hotel segments. The preprint appears on arXiv, an open-access repository that hosts scientific papers in fields including computer science and statistics before peer review [7]. As of late 2024, the platform was receiving roughly 24,000 new articles per month [7].