AI Citation Sources Study 2026-09
First round September 12–13, 2026; day-30 rerun September 25–26 | Each round: 36 questions × 2 engines × 3 samples = 216 runs
The question: when users in Taiwan ask about SEO and GEO in Traditional Chinese, which kinds of questions make AI assistants search the web, and what kinds of pages do they cite.
| Key finding | Number |
|---|---|
| Overlap between ChatGPT and Gemini citation sources | 17.8% (27 of 152 domains in common) |
| Share of "can you recommend a company" questions that triggered a search | ChatGPT 96.3%, Gemini 92.6% |
| Share of how-to and concept questions that triggered a search | ChatGPT 3.7%, Gemini 66.7% |
| Where recommendation answers drew citations | Job banks, e-commerce platforms and "best SEO agencies" lists written by agencies themselves |
| Where knowledge answers drew citations | Small personal blogs rather than high-authority sites |
The day-30 rerun used identical conditions. seobar.dev still did not appear in any citation, the overlap between the two engines was 14.7%, and Gemini's search rate fell for all three question types. Citation counts will not change meaningfully until a site has built visibility in search, so the series ended on day 30, and the raw data for both rounds is public.
What these numbers mean for SEO and GEO is covered in ChatGPT vs Gemini Citation Overlap: 17.8% in a 216-Sample Study.
About These Studies
Every study publishes three things: the questions or sample, the classification rules, and the raw data. The rules are written out because the same observations produce different numbers under different criteria, and without them nobody else can check the work or reuse it.
All data is released under CC BY 4.0. You are free to use and adapt it with attribution. If you get a different result from the same data, please let us know.