The Hotel List Position Becomes the Booking Clerk
The June 2026 arXiv paper Whose hotel does the AI recommend? An algorithm audit of reputation signals in LLM-assisted hotel selection, by Mirza Samad Ahmed Baig, Syeda Anshrah Gillani, and Asher Ali, turns hotel recommendation into a causal audit of what LLM assistants actually reward.
The Assistant Is the Shop Window
The paper, arXiv:2606.16344 [cs.AI], was submitted on June 15, 2026. Its object is ordinary and therefore important: a traveler asks an LLM assistant which hotel to book, and the assistant turns a set of properties into one recommendation.
That recommendation is not only advice. It is visibility. A conventional search page lets the user see alternatives, rankings, advertisements, snippets, and filters. A conversational assistant can collapse the visible market into a single named option plus a justification. The paper calls this an AI infomediary problem: the model stands between suppliers and customers, deciding which eligible property becomes legible.
This makes the study a useful companion to the site's work on recommender systems, AI search and answer engines, algorithmic transparency, brand reputation in answer engines, and search-agent recommendation reliability. The fresh angle here is not a hostile page. It is the apparently neutral presentation order inside a recommendation prompt.
What the Audit Tests
Baig, Gillani, and Ali run a pre-specified algorithm audit using a randomized choice-based conjoint. The assistant must choose among five synthetic hotel cards. The cards independently vary guest rating, review volume and recency, management response, chain affiliation, price, eco-certification, and list position.
The audit spans three traveler personas, nine prompt paraphrases, and twelve open-weight and proprietary models. The paper reports 3,024 main-arm choice sets per model and more than sixty thousand model calls across all arms. The full design, hypotheses, and analysis plan were specified and cryptographically hashed before confirmatory data collection.
The headline effects are legible. A top guest rating raises recommendation probability by 31.6 percentage points. A high price lowers it by 30.0 percentage points. Eco-certification adds 11.6 percentage points, and high review volume adds 8.3. A visible management response, despite being promoted in reputation-management advice, has no detectable effect in the pooled result at about 0.1 percentage points.
Position Is Not Empty
The most governance-relevant result is list position. The paper randomizes position independently of content, so position is not a proxy for quality. It is a content-free artifact of the candidate list. Even so, the first slot carries a causal advantage in the pooled analysis, worth about twelve dollars per night in the authors' price-equivalent framing.
That result matters because upstream systems decide order before the LLM speaks. A retrieval engine, booking platform, partner feed, advertiser interface, data broker, or prompt constructor can set the order of candidates. If the LLM then treats position as a recommendation signal, the upstream arranger has quietly become part of the booking clerk.
This is not the same as classic search-engine optimization. In search, a user can often see that something is first. In a chat recommendation, the position effect may disappear into the assistant's prose. The user sees a reasoned answer, not the fact that the answer inherited weight from an arbitrary card order.
Reasons Are Not the Weights
The paper also compares stated reasons with revealed weights. It reports positive but imperfect correspondence: models act on list position and review volume without naming those influences, while over-citing brand relative to its near-zero revealed influence.
That is the transparency problem in miniature. A model can produce plausible reasons that are not a faithful account of the causal weights that moved its recommendation. This does not require deception or intention. It is enough that the explanation layer is generated after the selection layer has already done its work.
For buyers, regulators, and suppliers, this means "the assistant explained itself" is not sufficient evidence. The relevant audit is behavioral: randomize inputs, measure shifts, compare stated reasons to revealed choices, and preserve the experimental record.
What It Does Not Prove
The audit does not model the entire travel funnel. Its limitation section says it isolates the assistant's selection among a fixed, already retrieved set of five candidates. It does not model retrieval, ranking, multi-turn negotiation, live booking inventory, advertising auctions, loyalty programs, or product-interface rules that might sit around a deployed assistant.
The hotel cards are synthetic. That is a strength for causal identification and a limit for ecological realism. The estimates show what moved model choice under controlled conditions; they do not certify any one travel platform's production behavior.
The paper also should not be read as a universal verdict on every model or every domain. It gives causal evidence for one high-value recommendation setting. The governance lesson is broader: when an LLM becomes an infomediary, interface artifacts can become economic signals.
Governance Standard
Any LLM-assisted booking or product-selection system should publish a recommendation audit card: candidate-source rules, ranking source, card order, paid-placement status, model and prompt version, attributes shown to the model, attributes hidden from the model, temperature and decoding settings, randomized signal tests, position-sensitivity tests, stated-versus-revealed reason checks, and appeal paths for suppliers.
For consumer protection, the key separation is between retrieval, ranking, and selection. A platform should say whether the model chose from an ordered list, whether order was randomized or controlled, whether paid or preferred partners entered first, and whether the explanation mentions factors that actually changed the recommendation.
The Spiralist rule is this: when a chat assistant recommends one option, the list position has already spoken. Audit the clerk before trusting the booking.
Sources
- Mirza Samad Ahmed Baig, Syeda Anshrah Gillani, and Asher Ali, Whose hotel does the AI recommend? An algorithm audit of reputation signals in LLM-assisted hotel selection, arXiv:2606.16344 [cs.AI], submitted June 15, 2026.
- arXiv experimental HTML for Whose hotel does the AI recommend? An algorithm audit of reputation signals in LLM-assisted hotel selection, reviewed June 25, 2026.
- Related pages: Recommender Systems, AI Search and Answer Engines, Algorithmic Transparency, The Brand Citation Layer Becomes the Reputation Map, The Planted Page Becomes the Recommendation Payload, and The Source ID Becomes the Factuality Test.