Sitegravity
← All research

Macau: ask the same question on another AI engine or in another language, and the answer changes

Published: November 9, 2026

In two Macau tests, the engine decided how many places an answer named, from none to a dozen. Language mattered for local food but hardly for resorts. Simplified and Traditional Chinese were closer to each other than to English for food, but still need to be read separately.

What we asked

Two Macau topics, asked the same way: food souvenirs (what to buy and where, 12 questions) and tourism and resort hotels (12 questions). Each topic has four questions per language, in English, Simplified Chinese and Traditional Chinese, covering the same four intents: discovery, recommendation, comparison and price. Every question was asked three times on each of four AI engines. In total: 144 answers per topic, 288 in all.

The questions are written in each language, not translated word for word, so the languages are comparable by intent but not by wording. This test did not include the Chinese engines Doubao and Yuanbao.

Main findings

  • The engine decided how many places an answer named. In the resort test, ChatGPT named no business at all in 21 of 36 answers (58%); DeepSeek and Gemini named at least one in every answer, on average 10.7 and 12.1 per answer against ChatGPT's 0.6. The 21 answers were not refusals: every one of the 36 ChatGPT answers used web search and cited sources, and the 21 gave general advice (price ranges, which area to stay in, how to split two days) without naming a hotel or shop. Of the 9 answers to price questions (3 languages × 3 rounds), ChatGPT gave this kind of answer in all 9 and Perplexity in 8, against 0 of 9 for DeepSeek and Gemini. In the souvenir test the gap was smaller: ChatGPT named no business in 7 of 36 answers (19%), DeepSeek in 0 of 36. For one resort we tracked, Gemini named it in 30 of 36 answers (83%), DeepSeek in 23 (64%), Perplexity in 9 (25%) and ChatGPT in none (0 of 36); none of the four resorts we tracked was named by ChatGPT in any of its 36 answers.
  • Language mattered for local food. One local bakery was named in 10 of 48 English answers (21%), against 27 of 48 in Simplified Chinese (56%) and 31 of 48 in Traditional Chinese (65%). English answers named no business at all in 9 of 48 cases, against 4 in Simplified and 2 in Traditional Chinese. For resorts the pattern was weaker: one resort was named in 9 of 48 English answers and in none of the 96 Chinese ones, while another showed the reverse (27 of 48 in Traditional Chinese, 18 in Simplified, 17 in English).
  • Simplified and Traditional Chinese were closer to each other, but still need to be read separately. Across the four bakeries we tracked, the average gap in mention rate between Simplified and Traditional Chinese was 7.3 percentage points, smaller than the gaps between English and each of them (14.1 and 18.2 points). In the resort test that did not hold: the three pairs were about equally far apart (9.9, 11.5 and 9.9 points), and one resort was named in 56% of Traditional Chinese answers against 38% in Simplified.

What it means

In Macau, "being visible to AI" is not one number. A business can be invisible to one engine and named in most answers by another, and can be named in Chinese but not in English, or the reverse. Check each engine and each language your customers use, and keep Simplified and Traditional Chinese separate.

Method and limits

Two tests on 31 October 2026, 144 answers each. Each language has four questions per topic with matched intents, so the differences between languages mix the language and the question wording. Rates are the share of answers in which an entity was named by our entity extraction; landmarks, temples and booking platforms are not counted as businesses. The businesses we tracked are the ones chosen for each test, not a ranking of the whole market.

Engine versions: we call the engines through their APIs, not the consumer apps, and the model names below are the ones recorded with each answer.

EngineModel used in the testsModel used now
ChatGPTopenai/gpt-6-luna through OpenRouter, native web searchgpt-6-luna, called directly through OpenAI
Geminigemini-3.1-flash-lite (souvenir test) and gemini-3.5-flash (resort test), through OpenRoutergemini-3.6-flash, called directly through Google
Perplexityperplexity/sonarperplexity/sonar
DeepSeekdeepseek-v4.1-flash through OpenRouter, with Exa web searchdeepseek-v4.1-flash

The names for ChatGPT, Perplexity and DeepSeek are unchanged, but providers can update a model behind the same name, so results today may differ. Gemini's model has changed, and the souvenir test used the smaller Lite version, which cited sources less often. The Chinese engines (Doubao, Yuanbao) are not in these data, so no claim is made about them. One test per topic, one date; models change.

Data collected on 31 October 2026. The businesses are not named.

Data collected October 31, 2026. Findings are anonymised and describe one test on one date. → Method

Want to see this for your brand?

Get a free AI visibility check