How do LLM hotel assistants weight reputation signals differently than humans?

A pre-specified algorithm audit runs a randomized, choice-based conjoint across twelve LLM travel assistants — four open-weight models run locally and eight proprietary models from OpenAI, Google, and Anthropic — each asked to recommend one of five synthetic hotel cards whose reputation attributes (guest rating, review volume, review recency, management response, chain affiliation, price, and eco-certification) and list position are independently randomized. Estimating the average marginal component effect (AMCE) of each attribute on the probability of recommendation, the audit finds rating and price dominate: "a top rating raises selection by 31.6 percentage points; a high price lowers it by 30.0." Benchmarked against the human electronic-word-of-mouth (eWOM) literature, where "review valence (rating) is consistently the dominant lever... with management response, brand, price, and green certification playing measured secondary roles," the assistants reproduce rating's primacy — rating "outranks review volume by nearly four to one" — but diverge in two ways: eco-certification is elevated to "the third most important signal (rivaling review volume)," while management response's effect "is statistically indistinguishable from zero" and brand affiliation "carries only a small penalty."

The paper's framing treats signaling theory and algorithmic gatekeeping as one system rather than two separate literatures: once an LLM assistant sits between the traveler and the property, "a second receiver is inserted into the signaling chain," so a reputation signal must be legible to, and weighted by, the machine before it can ever reach the human. The audit calls this the gatekeeper's "revealed reputation function," and concludes the assistant "partially honors and partially rewrites" the human hierarchy — it has "internalized the broad shape of the human signal hierarchy from its training distribution while operating, at the margins that matter commercially, on a reputation function of its own."

This gives causal, cross-model AMCE estimates for exactly the kind of cues that Do shopping agents respond more strongly to price and rating cues? shows LLM shopping agents over-respond to relative to a human baseline — rating and price again surface as the dominant levers, now quantified in percentage points rather than compared directly to a human control group. It also sits alongside Do language models favor sources regardless of item quality?: both find LLM selection swayed by attributes that are not the item's actual quality, but where that paper isolates a categorical identity cue (source domain) as either present or absent, this audit measures graded over- and under-weighting of attributes humans also use, against a documented human baseline ordering.

The audit isolates only the selection stage — choosing among a fixed, already-retrieved set of five hotels — and does not model retrieval, negotiation, or booking, so the estimates describe how the machine chooses among given options, not the full query-to-booking pipeline. The panel of twelve models is fixed-version and not a census of deployed assistants; the open-weight models are run quantized, which can shift output distributions from full-precision deployment. The excerpt also notes "stated reasons track revealed weights imperfectly" — the assistant's explanation for its choice does not reliably match what actually moved it. The evidence supports, but does not extend beyond, a selection-stage conclusion: a nascent "generative engine optimization" industry advising hotels how to win AI recommendations is, per the paper, "running well ahead of measurement," and signals hotels have rationally invested in for human audiences — management response, brand — may be repriced or ignored by the model that now gatekeeps demand.