The language of a question changes which sources AI cites
The same 24 questions, asked in English and in Vietnamese, drew on largely different sources. Four AI assistants also shared few sources with each other.
We asked four AI assistants the same 24 everyday questions in two languages and recorded every source each answer cited. The language of the question changed which sources the assistants relied on, and the assistants relied on different sources from each other.
What we tested
- Market: Vietnam. Questions about hotels, clinics, restaurants, education, professional services and shopping in Ho Chi Minh City, Hanoi and Nha Trang, plus two about Vietnam as a whole. No question named a business.
- Languages: English and Vietnamese.
- Assistants: ChatGPT, Claude, Perplexity and Gemini, through each company's developer access, with web search available and memory and personalization off.
- Repetitions: each question asked twice on each assistant in each language. 384 answers in total, all collected.
- Date: October 15, 2026.
A source is cited when the answer attaches it to a specific statement. We counted each cited website once per answer: 1,544 citations across 613 websites.
What we found
1. The language of the question changed the sources. For the same question, the share of websites cited in English that were also cited in Vietnamese was 36.8% on ChatGPT, 16.0% on Claude, 13.7% on Perplexity and 13.2% on Gemini. In Vietnamese, the assistants leaned more on sources based in Vietnam.
| Assistant | Sources based in Vietnam, English questions | Sources based in Vietnam, Vietnamese questions |
|---|---|---|
| ChatGPT | 64% (range 61 to 66%) | 75% (range 71 to 76%) |
| Claude | 57% (range 43 to 67%) | 94% (range 87 to 94%) |
| Perplexity | 32% (range 22 to 51%) | 85% (range 81 to 86%) |
| Gemini | 74% (range 67 to 76%) | 83% (range 79 to 83%) |
Some websites state no owner or location. The ranges show the result if all of those were counted one way, then the other.
2. The assistants relied on different sources from each other. For any two assistants, the websites both cited were between 8% and 16% of all the websites either one cited.
3. Where answers came from depended on the assistant. A business's own website was the most cited type of source on every assistant, but by very different margins. It made up about 79% of ChatGPT's citations in both languages, 52% to 59% of Claude's, 58% to 67% of Gemini's and 32% to 50% of Perplexity's. Perplexity drew more on booking sites, articles and directories.
4. Gemini often answered without searching. Gemini answered without using search in 64 of its 96 answers. Claude did so in 6 of 96. ChatGPT and Perplexity searched every time. Because Gemini searched so rarely, its figures above rest on fewer answers than the others.
5. Citations were spread thin. In most combinations of assistant and language, a website cited in just 2 of 48 answers was already among the 15 most cited.
The sources cited most often included trip.com, hotels.com, mytour.vn and pasgo.vn for travel and dining, michelin.com on ChatGPT, toplist.vn for Vietnamese questions on Perplexity and facebook.com on Gemini.
What this does not show
- Whether any cited source changed what an answer said.
- What the assistants' consumer apps show. Developer access can behave differently, and Gemini's rate of answering without search may differ in the app.
- Anything beyond these questions, these three cities, these assistants and this date.
- Whether the same pattern holds in other countries and languages. This is the first market measured with this method. The method is written down so it can be repeated anywhere.
The record behind this finding
- The questions, the classification rules and the reported numbers were fixed and committed before the first question was asked. Three dated amendments were added afterward, each before the step it governs.
- The Vietnamese questions were written by an AI model (Claude) and checked by a model from a different company (Gemini), with an independent back-translation. No native human reader checked them for this measurement. This matters most for the comparison between languages.
- Each cited website was classified by type and by whether it is based in Vietnam. Claude did the classification. Gemini independently classified a random 10% without seeing the first codes, and agreed on 93% of types and 95% of locations. Both are assistants measured here. Neither scored any answer, and neither was shown which assistant had cited a website.
- A business's own website is reported as a business website, never by name.
- All 384 answers are kept.
- Data: doi.org/10.5281/zenodo.23394836
Sandria takes no money from the AI platforms it measures.