Field note 01 · measured July to August 2026 · updated 27 October 2026

Same hotel, different universe.

We asked eight AI answer engines, four Western and four Chinese, the questions travellers actually ask about Phuket. The two halves named different winners. The Western engines searched the web as they answered; the Chinese models, queried without search, answered from what they learned in training. A Western-only audit sees half the picture.

What we measured

  • 494 live answers across two Phuket cohorts: luxury resorts (373 answers; 164 Western / 209 Chinese; 14 Jul to 1 Aug 2026) and wellness & longevity (121; 50 / 71; 14 Jul 2026).
  • 18 questions per cohort, drawn from our buyer-intent taxonomy: category, local, comparison and transactional, all asked in English. Questions naming any specific property were excluded from every figure here, so no brand is inflated by questions asked about it.
  • How each engine was queried: Perplexity, ChatGPT, Gemini and Claude with their web-search tools switched on, so their answers could draw on pages retrieved at question time (Claude supplied only 11 of the 214 Western answers). DeepSeek, Qwen, GLM and Kimi without web search: DeepSeek through its own API, which offers none, and Qwen, GLM and Kimi through a third-party host that serves them without it. Repeat-sampled; every answer stored verbatim.
  • Figures were computed under the same rules as the live Hotel AI Index, which has kept measuring since this window closed; scoring vocabulary is published at /methodology. This note is frozen to its measurement window.

Finding 1: the two halves crown different winners

In the wellness cohort, RAKxa Integrative Wellness appears in 32% of Western-engine answers (it was Gemini’s single most frequent first pick, eight times) yet reaches just 4% of Chinese-engine answers. The mirror image: Kamalaya is the Chinese engines’ favourite at 49%, double its Western share. Six Senses Yao Noi appeared in zero Western wellness answers while Chinese engines named it in 13%.

RAKxa Integrative Wellness32%4%Kamalaya24%49%Chiva-Som32%38%Six Senses Yao Noi0%13%Thanyapura Health & Sports2%10%

Western engines    Chinese engines · share of that side’s answers naming the property

The luxury cohort agrees at the very top, where Trisara and Amanpuri lead on both sides, then splits underneath: Rosewood Phuket nearly doubles its share on Chinese engines (27% → 48%), and COMO Point Yamu is five times more visible to Chinese engines than Western ones.

Trisara62%79%Amanpuri59%76%Rosewood Phuket27%48%Sri Panwa13%25%COMO Point Yamu2%11%

Western engines    Chinese engines · share of that side’s answers naming the property

Because the two halves were queried differently (see Finding 2), this data cannot say how much of each gap comes from the model and its training, and how much from searching versus remembering. It can say that a traveller would be handed a different shortlist.

Ask the West’s engines about wellness in Phuket and RAKxa keeps coming back first. Ask China’s, and it’s Kamalaya. Same question, different universe.

Finding 2: search versus memory

Across both cohorts, the Western engines searched as they answered and cited 2,087 sources over 214 answers, roughly ten per answer. The Chinese models, queried without search, cited none across 280 answers. That zero follows from how we queried them. It is not a discovery about the engines, and we would not want it read as one.

What the Chinese side does show is each model’s own picture of Phuket, with nothing looked up. That matters for two reasons. Not every route to these models searches: DeepSeek’s API offers no search, and the host we used served Qwen, GLM and Kimi without it, so any app or agent built the same way answers exactly as we measured. And memory is the slow lever: a page fixed this week can be retrieved next week, but what a model learned about a property changes only when the model is retrained.

2,087sources cited · 214 Western answers · search on
280answers from memory · Chinese models · search off

Scope, stated plainly: search was on for the Western four and off for the Chinese four, so this note compares search and memory as much as West and China, and its data cannot separate the two. Consumer apps such as Doubao, Yuanbao, DeepSeek and Qwen can search the web in some modes. These figures describe the models as we queried them, not every product built on them.

Finding 3: where the Western answers come from

The Western engines’ wellness answers leaned on a small, concrete set of sources: TripAdvisor, World Spa Awards, Healing Holidays, BookRetreats, health.travel, specialist press like The Hotel Journal, and the properties’ own sites (chivasom.com, rakxawellness.com). Gemini additionally wraps most of its citations in its own grounding redirector, which we count but do not name as a source.

That list is an instruction: it is the earnable surface. Presence on those exact domains is what a Western engine can retrieve and cite next quarter.

What we think it means

This section is interpretation; everything above is measurement.

  • Search and memory are two different games. Retrieval is what a search step finds and cites: earn presence on the sources engines actually read, and answers can change within a quarter. Memory is what a model already believes: it moves over training cycles and rewards sustained presence in the sources models learn from, which for models trained on large amounts of Chinese-language text means travel platforms, encyclopedic entries, booking listings and editorial coverage in Chinese.
  • A Western-only audit leaves part of the exposure unmeasured. If mainland China, Hong Kong, Taiwan or Chinese-speaking Singapore are a meaningful share of a hotel’s arrivals, the answers these models give are part of its AI exposure, and a Western-only audit never sees them.
  • Divergence is opportunity, in both directions. RAKxa (32% Western, 4% Chinese) is found by search but barely known to the Chinese models: a memory gap, slow to close. Kamalaya (24% Western, 49% Chinese) is well known to those models but less prominent in what Western search surfaces: a retrieval gap, closable this quarter. Neither shows up in a single blended score, which is why we report every property per engine, per side.

Limits

One destination, two cohorts, 494 answers, one measurement window. Engines update continuously, and these figures will drift. All questions were asked in English; a traveller asking in Chinese may get different answers. Repeat sampling narrows but does not eliminate answer variance. And AI visibility is measured presence in answers, not bookings: we make no revenue claims here. The quarterly Index is where movement gets tracked.

Where does your property sit in each universe?

The audit runs your guests’ real questions across all eight engines and shows both halves side by side: verbatim, scored and benchmarked.

Get your free audit