Same question, another language: about half of AI's recommendations change
Published: November 3, 2026
We asked the same realistic questions in four or five languages on four AI engines, across four Malaysian industries. The recommended lists were about half the same: from 21% to 63% depending on the industry.
What we asked
We wrote realistic customer questions, each with a place, a budget and a situation, once for every language. The same scenario was asked in English, Simplified Chinese, Traditional Chinese and Malay (and Indonesian for private hospitals). Every scenario was asked three times on each of four AI engines: ChatGPT, Gemini, Perplexity and DeepSeek.
We ran this for four industries in Malaysia: private universities, attractions and theme parks, private hospitals, and restaurants. In total we collected 612 answers.
Main findings
- About half of the recommended businesses differ between languages. Across the four industries, the median overlap between the recommended lists of two languages was 50% (84 language pairs, range 0% to 100%).
- It depends on the industry. The lists agreed most for private hospitals and least for restaurants:
| Industry | Language pairs | Median overlap | Range |
|---|---|---|---|
| Private hospitals | 30 | 63% | 38%–89% |
| Private universities | 18 | 56% | 21%–100% |
| Attractions and theme parks | 18 | 40% | 33%–64% |
| Restaurants | 18 | 21% | 0%–46% |
- The core of a list stays; the edges move. In the first three industries a small group of businesses was recommended in every language, while others appeared in only one or two. Restaurant answers were the most scattered, so they had almost no shared core.
- The engines disagree as much as the languages do. In three of the four industries, DeepSeek and Gemini named the business we tracked far more often than ChatGPT and Perplexity did:
| Engine | University | Attraction | Hospital | Restaurant |
|---|---|---|---|---|
| DeepSeek | 81% | 8% | 64% | 14% |
| Gemini | 50% | 0% | 67% | 8% |
| Perplexity | 17% | 11% | 24% | 0% |
| ChatGPT | 0% | 0% | 22% | 0% |
Each figure is the share of answers that named the one business we tracked in that industry.
What it means
- A result in English describes only part of what customers are told. Customers asking in Chinese or Malay can be given a noticeably different list.
- An average across engines hides large gaps. A business can be named by one engine in most answers and by another in none.
- Where answers are very scattered, as with restaurants, being named at all is the first goal.
Method and limits
Each language and scenario has 12 answers in these lite tests, so differences of a few points are noise. The overlap counts only businesses the AI recommended and that appeared in at least 2 answers for that scenario and language. These are four tests on one date, in one country; AI models change often. The definitions and limits are on the Method page.
Data collected on 1 and 3 November 2026. All findings are anonymised.