Hotels Say Friendly. AI Says Upscale.

There has been a lot of talk about how smart AI engines are. Underneath, they started life as large language models, and language is still how they work: a guest asks in a sentence, the engine answers in sentences, and the hotels it recommends arrive wrapped in words. So we looked at the words.

This is a much different exercise from a post we originally published several years ago, before the first version of ChatGPT had even arrived: Choosing the Best Keywords for Hotels. Inside AI engines, tracking keywords or even phrases tells you very little; people now talk to machines in natural language and the machines answer in kind. But descriptive language still matters, and in our opinion it matters more in hospitality than most verticals. A stay is an experience, and the words used to describe it are part of what a guest is choosing.

Our on-going Boutique Hotel AI Index asks five AI platforms (ChatGPT, Gemini, Copilot, Google AI Mode and Google AI Overviews) for boutique, luxury and romantic hotel recommendations in 100 destinations worldwide. We chose those three words from our own work. Each one describes both a type of hotel and the experience a guest expects to have there, and all three are common currency in the industry and in everyday speech. Everyone carries their own idea of what each word (and experience) means, which is exactly why how AI engines interpret them is a fascinating dive into adjectives, nouns and all sorts of grammatical structures.

The analysis below is based on some 38,900 full-text answers. We analysed every answer to find the descriptive words, then counted the share of answers each word appears in, once per answer however often it repeats. Words that praise rather than describe (top, good, perfect) and words that name a feature rather than a quality (private, free) were set aside; the full list is in the data caveats.

The Question Sets the Vocabulary

Let’s start with the vocabulary our three prompt sets share. Three words sit in about three quarters of all answers: historic (74%), intimate (73%) and modern (73%). Historic and modern mean opposite things, yet about half of all answers use both and nine in ten use at least one. Historic describes the fabric: historic charm (15% of answers), historic architecture and historic building (9% each). Modern describes what is inside it: modern luxury (21%), modern design (16%). An old shell with new contents is very much a house formula of many AI hotel answers.

Share of each question’s answers containing the word.

Even the shared core leans with the question. Intimate is in 78% of boutique answers and 89% of romantic ones, but only 53% of luxury answers, where classic and contemporary come first. Romantic turns up in 55% of boutique answers and 47% of luxury ones, so the romantic register spills well beyond the romantic question.

Below the core, the three vocabularies part company. Measured as a word’s share within one question against its share across all answers:

Boutique over-indexes on retro, independent, industrial-chic, creative, eclectic, family-run, quirky, distinctive, vintage and artsy, each at roughly twice the overall rate. It rarely says timeless, world-class or grand.

Luxury over-indexes on prestigious, impeccable, premier, expansive, top-tier and five-star, and, less obviously, family-friendly at twice the overall rate. It says cozy, quirky and charming at a third of their overall rate or less.

Romantic over-indexes on magical (about twice the overall rate), unforgettable, indulgent, breathtaking, cozy, secret, scenic and secluded. Unsurprisingly, it almost never says family-friendly.

Treat these lists as food for thought: a picture of the language a guest now meets when they ask AI engines for hotel recommendations. They may be useful when writing copy or simply when picturing how a hotel is being presented, but they are certainly not meant to be interpreted as a check and forget keyword list for your content team.

Hotels Say Friendly, AI Says Upscale

Studying the homepages of the hotels named in the study (2,916 English-language homepages of sites that serve a single hotel) provided an illuminating comparison. A homepage describes one hotel to a prospective guest; an answer describes several hotels to someone choosing between them. This comparison showed how the two registers differ.

Of the twenty descriptive words each side uses most, they share five: modern, historic, elegant, intimate and contemporary. Across the stricter and looser definitions of a descriptive word we tried, the count ran from four to seven.

Share of homepages, and of answers, containing the word.

Read down the hotel side and two kinds of word dominate. One is about people: friendly, helpful, professional, thoughtful, genuine. On homepages these attach to staff, service, team and hospitality; “friendly staff” alone is on 5% of homepages. The other is about how the stay will feel: unforgettable, memorable, delightful, with “unforgettable experience” on 9% of homepages.

Hotels describe how you will be looked after and how you will feel while staying there. The engines describe what kind of hotel it is: a positioning (upscale, chic, sleek, design-forward) and a price tier (high-end, ultra-luxury). “Vibe” is in half of all AI answers and on 6% of homepages. Much of that is Gemini, which gives nearly every hotel it lists a “Vibe” line of its own; the other engines use the word to sort the options (“depending on the vibe you are after”) or to sum a hotel up (“a relaxed, upscale vibe”).

This describes two different vocabularies; it is not a list of words to put on your website. What it does show is that the engines are not echoing hotels’ own language. They summarise hotels in a register of their own, and that register moves with the question, the platform and the format of the week.

Each AI Engine Has its Own House Writing Style

Each engine answered about 8,400 of the questions across the 28 runs, except Google AI Overviews, which only appeared for 5,297. The figures are the share of that engine’s answers against the share of all answers.

ChatGPT says outstanding (20% against 5%), memorable (23% against 6%) and special (47% against 16%). It writes for first-time visitors (27% against 6%), and describes hotels as small (44% against 17%).

Gemini says massive (39% against 13%), incredible (44% against 17%), gorgeous (36% against 15%), moody (17% against 6%) and jaw-dropping.

Copilot says limited (50% against 12%), premium (72% against 20%), budget-friendly (29% against 7%), affordable, expensive, busy, noisy and remote: the words of a price column and a trade-offs section, which is what Copilot’s answers are built around.

Google AI Mode says posh (31% against 8%), upmarket (26% against 6%), plush, airy, luxe, casual and swanky.

Google AI Overviews says celebrated, highly-rated and tailored, each at between 1.7 and 2.2 times its overall rate, and otherwise says little the others do not.

Only one of the engines regularly tells you what might be wrong. Copilot mentions a downside (something limited, busy, expensive, noisy, dated or crowded, or an explicit trade-off) in between 84% and 99% of its answers. It has a “Risks & Trade-offs” section and uses it.

On the latest run, 19 October, Google AI Mode mentioned a downside in 6% of answers, AI Overviews in 4% and Gemini in 11%.  AI Mode and Gemini both said more in early summer: AI Mode in between a sixth and a third of answers through July, Gemini in 10% to 17% on all but one run until early August (33% on 29 June).

ChatGPT went the other way: between 9% and 26% of answers mentioned some sort of downside until 5 October, then 47% on 12 October, and 55% on 19 October (also the dates a model change was noted in our monitoring tool). This change in answer style made two blocks much more common in ChatGPT answers, the “Trade-off’ and ‘Keep in mind’ sections we discussed on LinkedIn here.

The Words Used Change in Steps

One of the earliest findings in our study was that AI hotel recommendations churn. If you ask the same AI engine for hotel recommendations an hour apart the same number one hotel is named only 60.2% of the time.

The vocabulary used behaves differently. Asked twice on the same day it barely moves: a median of one to two points per word on the four engines whose repeat data is complete. Then, on a single run, it jumps.

On Google AI Mode, “elegant” went from 82% of answers to 9% between the runs of 27 July and 3 August, “upscale” from 89% to 30% and “posh” from 58% to 2%; these descriptors had largely left the answer format.

On ChatGPT, “elegant” fell from 77% to 48% on the run in August where the platform’s model label changed, and rose from 49% to 84% on the run in October where it changed again, holding at 83% the following week.

Gemini took up the posh-and-upscale vocabulary in the first week of September and had largely dropped it again by mid October.

Copilot’s vocabulary is the steadiest of the five across all 28 runs, apart from the 13 July run, when half of its answers came back without a hotel list.

We state these as associations with format and model changes for clarity. We cannot see inside the platforms and it is not provable if the change in models used results in vocabulary changes. What we can say with some certainty is that the words an AI platform puts around a hotel are partly a product of the platform itself.

They are not a transcript of what the hotel says about itself, and they are not a fixed description of it either. A hotel that sees the adjectives around its name change from one month to the next has not necessarily changed; the engine may have. Our suggestion is to read the words as a description of the describer as much as of the described.