Findings
- Vendor pages are the largest source. In all three engines and both languages they come first, 56.9% overall.
- GitHub is not a niche source. As one domain, GitHub is 14.4% of all citations. About seven in ten Perplexity answers (68.3% in Chinese, 70.0% in English) cite at least one GitHub or package page.
- Official docs are a tenth, and it depends on the engine. 22.3% for ChatGPT in English, 3.9% for Gemini in English, with ranges that do not overlap.
- Chinese questions lean on tech communities. Gemini 27.0% in Chinese and 6.6% in English; Perplexity 13.9% and 3.3%. ChatGPT is low in both, with no visible difference.
- Video is hardly cited. 20 of 5,456 citations, 0.4%; ChatGPT cited none.
- Sources do not hold still. Asked again five days later, the same question shared only 25–39% of its domains across the three engines. A single run says little about “who AI cites”.
Source types by engine
Each cell is that row’s share of all citations, both runs together.
| Engine | Language | Answers | Citations | Per answer | Official docs | GitHub and packages | Vendor pages | Directories | Tech communities | Video | Other |
|---|---|---|---|---|---|---|---|---|---|---|---|
| ChatGPT | Chinese | 120 | 635 | 5.3 | 14.8% | 20.0% | 58.9% | 3.8% | 1.6% | 0% | 0.9% |
| ChatGPT | English | 116 | 762 | 6.6 | 22.3% | 8.9% | 64.3% | 1.2% | 3.0% | 0% | 0.3% |
| Perplexity | Chinese | 120 | 1,200 | 10.0 | 10.7% | 18.4% | 50.0% | 4.9% | 13.9% | 0.3% | 1.8% |
| Perplexity | English | 120 | 1,197 | 10.0 | 6.3% | 21.8% | 60.2% | 7.9% | 3.3% | 0.4% | 0.1% |
| Gemini | Chinese | 119 | 795 | 6.7 | 10.8% | 14.7% | 34.3% | 10.4% | 27.0% | 0.4% | 2.3% |
| Gemini | English | 115 | 867 | 7.5 | 3.9% | 5.3% | 74.3% | 8.3% | 6.6% | 0.9% | 0.7% |
| All | Chinese | 359 | 2,630 | 7.3 | 11.7% | 17.7% | 47.4% | 6.3% | 14.9% | 0.3% | 1.7% |
| All | English | 351 | 2,826 | 8.1 | 9.9% | 13.3% | 65.6% | 6.2% | 4.2% | 0.5% | 0.3% |
The 10-29 run asked each question 3 times, so it weighs more in these totals. Perplexity gave almost the same sources when asked again in one run (see stability below), so its 90 answers there are closer to 30 samples.
How reliable the differences are. Each row rests on 30 questions. Resampling the questions 2,000 times gives these 95% ranges:
| Comparison | One side | Other side | Ranges overlap? |
|---|---|---|---|
| Official docs: ChatGPT English vs Gemini English | 14.6–31.1% | 1.8–6.3% | No |
| GitHub: Perplexity English vs Gemini English | 14.6–29.5% | 2.7–8.6% | No |
| Vendor pages: Gemini English vs Gemini Chinese | 66.4–81.8% | 26.6–42.6% | No |
| Tech communities: Gemini Chinese vs Gemini English | 21.1–32.7% | 3.7–9.8% | No |
| Tech communities: Perplexity Chinese vs English | 9.9–18.2% | 1.3–6.0% | No |
| Tech communities: ChatGPT Chinese vs English | 0.3–3.2% | 1.3–5.1% | Yes, no visible difference |
| Official docs: ChatGPT Chinese vs English | 8.9–20.8% | 14.6–31.1% | Yes, no visible difference |
How many answers cite at least one
These shares count answers, not citations.
| Engine | Language | Official docs | GitHub and packages | Vendor pages | Directories | Tech communities | Video |
|---|---|---|---|---|---|---|---|
| ChatGPT | Chinese | 34.2% | 35.0% | 73.3% | 15.0% | 7.5% | 0% |
| ChatGPT | English | 54.3% | 29.3% | 83.6% | 6.0% | 15.5% | 0% |
| Perplexity | Chinese | 42.5% | 68.3% | 90.0% | 29.2% | 65.8% | 3.3% |
| Perplexity | English | 27.5% | 70.0% | 100% | 40.0% | 24.2% | 4.2% |
| Gemini | Chinese | 27.7% | 54.6% | 74.8% | 42.9% | 73.1% | 2.5% |
| Gemini | English | 21.7% | 26.1% | 97.4% | 35.7% | 32.2% | 6.1% |
How stable the sources are
Overlap = domains cited both times ÷ domains cited either time, averaged over questions. “Same run” compares the 3 answers to a question on 10-29, pair by pair, sent almost together (a median of under a minute apart). “Five days later” compares the first of them with the 11-03 answer.
| Engine | Language | Same run | Five days later |
|---|---|---|---|
| ChatGPT | Chinese | 34.9% | 32.6% |
| ChatGPT | English | 35.6% | 30.7% |
| Perplexity | Chinese | 99.3% | 34.3% |
| Perplexity | English | 96.9% | 38.6% |
| Gemini | Chinese | 36.6% | 30.8% |
| Gemini | English | 31.2% | 24.9% |
ChatGPT and Gemini share only a third of their sources even when asked again straight away. Perplexity repeats itself in the short term and changes within days.
Split by run, the order of the main types barely changes. The largest move is Perplexity in Chinese: GitHub and packages from 15.7% to 26.7%, tech communities from 15.6% to 9.0%.
Also found
- Sources are scattered. 872 domains in all, 469 of them seen once or twice. The top 10 domains are 29.1% of citations, and GitHub alone is 14.4%.
- Most official docs are the platforms’ own. Of 587 official citations, 403 are platform developer pages and 184 are AI tools’ own docs, of which 182 are on questions that name an AI client.
- English answers cite lists and comparisons more. Citations whose title has words like best, top N, vs, alternative, 推荐 or 对比 are 15.9% on English questions and 6.2% on Chinese ones, highest for Gemini in English at 21.9%. This is a rough rule on titles.
- Reddit itself is cited little. reddit.com is 1.1% of all citations, highest for Gemini in English at 3.8%.
Method
- What was asked. 30 questions each in Chinese and English on how to get social media data, naming no data vendor, such as “YouTube transcript API for LLM apps”, “Douyin API to get video comments”, “小红书笔记搜索 API 推荐” and “微博热搜 API”. Some name an AI client (how to let one read Xiaohongshu, for instance).
- Whom and how. ChatGPT search, Perplexity and Gemini search, through a third-party service that collects AI answers, with no signed-in user.
- How often. The 10-29 run (finished between UTC 10-29 and 10-30) asked each question 3 times and the 11-03 run once: 720 requests, 10 failed, 710 answers. Of these, 38 listed no sources; they count as answers and add no citations.
- What counts as a citation. The source list of the answer for ChatGPT and Gemini, and the sources attached to the answer for Perplexity. A page (address without its parameters) counts once per answer.
- How sources were sorted. By domain:
- Official docs: social platforms’ developer and help pages, and AI tools’ own docs.
- GitHub and packages: GitHub, code hosting and package registries.
- Vendor pages: sites of companies offering data or tools, with their docs and blogs.
- Directories: MCP and API directories, app and extension stores, product review sites.
- Tech communities: developer blogging platforms, cloud companies’ developer communities, forums, Q&A sites, personal blogs and posts on social platforms.
- Video: YouTube and Bilibili.
- Other: news media, academic pages and pages that could not be judged.
- A domain on no list counts as a vendor page.
Limits
- Few runs per question. The 11-03 run asked each question once. The ranges above reflect the choice of questions, not time, and the sources themselves change fast.
- We chose the questions. All 60 are about social media data and lean towards developers and AI tool users. They do not stand for AI search as a whole, and other questions would give other shares.
- Sorting by domain is coarse. A cloud company’s domain holds both articles and product docs and counts as a tech community; a vendor’s blog and docs both count as vendor pages.
- The long tail defaults to vendor pages. Domains seen once or twice and on no list are 8.4% of citations. Some may be personal blogs or media, so the vendor share may run slightly high.
- Not what a real user sees. Answers came through a third-party service, with no sign-in, region or chat history, and may differ from what you get on the web.
- A citation is not adoption. Being listed does not mean the answer used the page’s wording or that anyone opened it.
- We are not bystanders. This site is one of the vendors in this space, and its pages count as vendor pages. This report gives no figure for any single vendor and no rank of our own.
What this is good for
- On questions like these, vendors’ product pages and docs are the type cited most. Sources are spread over 872 domains, 469 of them seen once or twice, so one page on a small site can get in.
- For open source or code-based tools, treat the GitHub repository as a real page. This shows most on Perplexity.
- On Chinese questions, articles in tech communities are what gets cited, above all on Gemini. This does not hold on English questions.
- Video is hardly cited on questions like these, so do not count on it to get into such answers.
- One run is not enough. Ask each question several times, again days later, and read the shares and ranges, not a single “was I there”.