THINK BIG
GEO Source Analysis · Local Brand Placement
Who Does ChatGPT Cite When It Answers Questions About Taiwan? — Where Local Brands Should Put Their Information
15,729 citations reveal: ChatGPT leans on brand sites and government bodies, roughly 70% not in Chinese. Perplexity leans toward social and video. Three AIs, three separate playing fields — local brands need a separate plan for each.
1. What's happening: same question in Chinese, three completely different source lists
In July 2026, an independent study in Taiwan published a rare dataset. The researcher wrote 60 consumer-decision questions in Traditional Chinese, sent them through the APIs of three AI services (ChatGPT, Gemini, and Perplexity), repeated each question multiple times, and collected 15,729 citations, each tagged by source type and language. (Source: Taiwan Traditional Chinese AI citation study, July 2026)
Source types cited by each AI (selected)
| Source type | ChatGPT | Gemini | Perplexity |
|---|---|---|---|
| Brand websites | 55.8% | 35.5% | 18.6% |
| Government / research institutions | 21.2% | 1.3% | 0.7% |
| Review and comparison sites | 6.1% | 20.5% | 21.4% |
| News media | 1.4% | 12.0% | 8.8% |
| Forums | 7.4% | 3.2% | 10.8% |
| Social media | 0.1% | 0.1% | 7.9% |
| Video | 0.0% | 1.4% | 6.7% |
Language of cited content
| Language | ChatGPT | Gemini | Perplexity |
|---|---|---|---|
| Traditional Chinese (Taiwan) | 30.1% | 71.4% | 87.8% |
| English | 44.4% | 18.1% | 5.5% |
| Japanese | 23.1% | 4.3% | 1.5% |
Two more findings stand out: ChatGPT doesn’t always search. It triggered a web search on only 46% of questions and answered the rest from model memory with no sources. For brand-comparison questions, though, the trigger rate rose to 70%. Three AIs, three separate playing fields. ChatGPT’s cited domains overlapped only 5.4% with Perplexity’s and 6.3% with Gemini’s. Perplexity and Gemini overlapped just 20.1%.
Read these numbers with their limits in mind. This is independent, non-peer-reviewed research. It used APIs, not the consumer interface. It ran in a single time window in July 2026. Questions were consumer decisions; the researcher notes ChatGPT frequently cited official U.S. and Japanese government sites, so results for business-services questions may differ. Use it for direction, not as a law for every industry.
2. What the sources say: why does ChatGPT lean so hard on official sources?
Other research points the same way. An April 2026 study across ChatGPT, Google AI Overviews/Gemini, and Perplexity analyzed 602 prompts and more than 21,000 citations. It found that Perplexity and Google cite more sources, more broadly, while ChatGPT cites fewer but each page it reads carries more influence on the answer. High-impact sources tended to be longer, well structured, and closely relevant to the question. (arXiv 2604.25707)
Whether ChatGPT searches determines whether you can influence it. A 2026 University of St. Gallen study found that 57.8% of ChatGPT’s runs had no citations at all, because it triggers web search only for certain queries. (arXiv 2604.07585) When it doesn’t search, it relies on what it learned in training — and what shapes those answers is consistent public information across many sites over time, not the web page you updated yesterday.
What the platforms say. OpenAI: ChatGPT search builds its index with OAI-SearchBot, and sites that block it won’t appear in ChatGPT search answers. (OpenAI crawler overview, Help Center) Google: AI Overviews and AI Mode draw on Google’s own index; one of Google’s recommended practices is keeping Merchant Center and Google Business Profile information up to date. (Google Search Central) Perplexity: its index is built by PerplexityBot, which Perplexity says is not used to crawl content for training AI models. (Perplexity crawler docs)
Forum buzz can’t be manufactured. Perplexity does cite forums and social media, but paying for forum recommendations is a dead end. Dcard permanently bans accounts used for word-of-mouth marketing or brand promotion, and an undisclosed paid endorsement in a consumer’s voice may violate Taiwan Fair Trade Commission rules on endorsement advertising. (Dcard rules, Taiwan FTC guidelines) What does work: posting genuinely useful content from your own named accounts.
3. What we changed after reading the study
Our business is in Taiwan, and we’ve always published primarily in Chinese. The biggest takeaway for us: if most of what ChatGPT cites for Taiwan questions isn’t in Chinese, a Chinese-only website starts at a disadvantage there.
Here’s what we did:
1. An English edition of every NEWS article. Starting this week, every article ships with an English version, and the two pages are linked with hreflang so AI systems know they cover the same topic. The English version isn’t a word-for-word translation; it’s rewritten for English-speaking readers.
2. One consistent official identity. Our legal name, English name, business registration number, location, and service description now read the same across our website, social bios, and listings. Our company registration with Taiwan’s Ministry of Economic Affairs is already public government data, and the description on our website matches it.
3. A separate plan per AI, not one-size-fits-all. Our website and English pages for ChatGPT; site structure and Google Business Profile for Gemini and Google AI Mode; long-form posts on our own named social accounts for Perplexity.
4. Still in progress: listings in industry association and public directories, and bylined media coverage. We’re building these from a low starting point, and we won’t use fake accounts or paid posts to fill the gap faster.
An AI once told us, with total confidence, that our company had nothing to do with AI. Looking back, it most likely couldn’t find enough consistent information about us and pieced together fragments. Making our identity consistent across multiple trustworthy places is, in our view, the first thing worth doing.
4. Our recommendations: a per-AI placement plan for local brands
Start with which AI your customers use, then decide where to put your information. The table below reflects the study’s observations; treat it as direction, and measure your own industry.
| To get mentioned by | It tends to cite | Where to put your information |
|---|---|---|
| ChatGPT | Brand websites, government and research institutions; high share of English | A bilingual website with Chinese and English pages linked to each other; government and industry association directories; institutional-style content such as method write-ups and white papers |
| Gemini / Google AI Mode | Brand websites, review sites, news media | A website Google can index and show snippets for; Google Business Profile; bylined media coverage and review-site listings |
| Perplexity | Review sites, brand websites, forums and social, video | Long-form posts from your own named social accounts (e.g., Threads); YouTube videos with transcripts; third-party comparison lists |
Four things that help with every AI:
1. Keep your name, address, phone number, and registration number consistent. Decide how your Chinese name, English name, and short name are written, and use them everywhere. 2. Use Taiwanese terminology and Taiwan-specific details. Traditional Chinese, local wording, a full address, a +886 phone number, and your registration number all tell AI you’re a Taiwanese business. 3. Publish brand-comparison content. ChatGPT searches most often on brand-comparison questions, so neutral “A vs. B” pages have the best chance of being read. Stick to facts. 4. Put important information in text on the page. Not only in images, PDFs, or collapsed tabs.
Bottom line: not every AI reads the same sources. Find out which AI your customers ask, then put accurate information where that AI looks.
Measure first. Three AIs, three plans. Compliant channels only.
(The following reflects Think BIG’s own position and is separate from the neutral reporting above.)
In our search visibility service (SEO + GEO), recommendations are always written separately for each AI: Measure first, then place — we use repeated sampling to see where you stand on each AI and which sources each one cites, then decide where to invest first. Three AIs, three plans — ChatGPT: your website (including English pages) and institutional directories. Gemini and Google AI Mode: site structure, Google Business Profile, and bylined coverage. Perplexity: named social accounts and video.
Consistent identity is the foundation. We align your company name, description, and contact details everywhere, so AI can recognize you and describe you correctly. Compliant channels only — no paid forum posts, no fake accounts, no fake reviews; paid partnerships are always disclosed.
We don’t promise rankings or sales results. We promise to put the right information in the right places and show you the numbers every month.
FAQ
- Q1: What sources does ChatGPT cite most when answering questions about Taiwan?
- According to a July 2026 Taiwan study (60 consumer-decision questions, 15,729 citations), 55.8% of ChatGPT’s citations were brand websites and 21.2% were government or research institutions, about 77% combined. Blogs and news media made up very small shares. It’s an independent, single-window, API-based study, and results may vary by industry.
- Q2: Why does ChatGPT cite English sources when the question is in Traditional Chinese?
- In the same study, 44.4% of ChatGPT’s cited content was in English and 23.1% in Japanese, with only 30.1% in Traditional Chinese. The researcher observed frequent citations of official U.S. and Japanese government sites. For Taiwanese brands, a consistent English version of your website is an additional way to be read by ChatGPT.
- Q3: Will posting recommendations about myself on Dcard or PTT improve AI visibility?
- We don’t recommend it. Dcard permanently bans accounts used for word-of-mouth marketing or brand promotion, and an undisclosed paid endorsement in a consumer’s voice may violate Taiwan Fair Trade Commission rules. A better approach is sharing genuine, useful content from your own named accounts.
- Q4: Do different AI services really cite different sources?
- Yes, very different ones. The same study found ChatGPT’s cited domains overlapped only 5.4% with Perplexity’s and 6.3% with Gemini’s, and Perplexity and Gemini overlapped just 20.1%. Being visible on one AI doesn’t mean you’re visible on another, so placement and measurement both need to be done separately.
- Q5: What should a local small business do first?
- Make your company identity consistent online: Chinese and English names, address, phone number, registration number, and service description should read the same on your website, Google Business Profile, social bios, and public directories. It’s low-cost and helps both Google Search and every AI recognize you.
This article combines research reporting with Think BIG’s first-hand practice. All cited studies and official documents are linked; research figures reflect the time period, platforms, question types, and test conditions stated in each source. Our own changes are described only in terms of direction and relative progress; no absolute figures are disclosed. Nothing here is directed at any company or product.