Can a hotel ever see where it ranks in an AI travel assistant's shortlist, or push back on that ranking?

This explores whether hotels can see, or push back on, where they sit in the candidate list an AI travel assistant works from. The short answer from the corpus is no, and that is the problem. A randomized audit of twelve LLM travel assistants held every quality attribute fixed and moved only a hotel's position in the list. Position alone changed which hotel got picked, and the effect was worth about twelve dollars a night, comparable to real reputation signals Does list order alone shift which hotel an LLM recommends?. The authors call it a content-free lever: nothing about the hotel causes it, and the hotel has no window into it. The order is set upstream by whatever system collects the candidates, often a retrieval or memory layer kept separate from the model's reasoning How can LLM agents handle huge candidate lists without breaking?. That layer is invisible to the businesses being ranked.

What hotels can influence is the content the model reads, and the same audit team found the model doesn't weigh content the way guests do. LLM assistants match humans in treating rating and price as most important. But they give eco-certification about as much weight as review volume, and they give no measurable weight to management responses to reviews How do LLM hotel assistants weight reputation signals differently than humans?. So a hotel that carefully answers every review is spending effort on a signal the AI ignores, and a certification badge may count for more than it would with a human guest. The fight that can actually be won happens over attributes, not position.

The broader research on LLM judges suggests the hotel case is one example of a general pattern. AI evaluators reward things that have nothing to do with quality: fake references and rich formatting raise scores without any access to the model Can LLM judges be tricked without accessing their internals?. In hiring, models prefer resumes written in their own style, and applicants who used the evaluating model were shortlisted 23 to 60 percent more often Do language models favor resumes they rewrote themselves? Do LLM evaluators favor resumes written by their own model?. There is early evidence that this happens because models recognize their own writing Do LLMs favor their own text because they recognize it?. A hotel could plausibly write its listing in the assistant's own style. That would be gaming, though, not contesting, and the corpus doesn't test it for hotels. The darker version already has a name: advertisement embedding attacks, where promotional content is slipped into the model or its distribution channel while accuracy looks untouched Can language models be hijacked to embed hidden advertisements?.

If individual hotels can't contest their position, the realistic remedies sit with whoever runs the system. Evaluation research points to two ideas. First, aggregating many evaluations, especially across different models, pulls rankings toward expert consensus Can aggregated LLM rankings match human expert judgments?. Applied here, shuffling the candidate order and averaging over several runs should wash out position effects, though no paper in the corpus tests that on hotel lists. Second, planted test cases can act as alarms for judge failure Can deterministic checks protect LLM judges from failure?, and that is essentially what the hotel audit did from outside. The surprise is that the most practical way for a hotel to 'see' its position may be to run that kind of audit itself, since the platforms don't expose the list.