We read every source in 479 AI answers to the same 136 buying questions in Korea and the US. 88% of the websites cited for a question appeared on one engine only, Perplexity cited Reddit in 34% of US answers and Naver AI Tab cited Naver Blog in 97%.
The data
This analysis reads the sources of every answer in the AI Index run of September 30, 2026: 14 categories of buying questions, seven in South Korea and seven in the United States, 136 questions and 479 answers. Korean questions went to ChatGPT, Gemini, Perplexity and Naver AI Tab, and US questions to ChatGPT, Gemini and Perplexity, each from inside its market, one answer per question and engine. The brand rankings from the same run are in Which brands do AI assistants recommend?
A source is a link the answer lists as a citation. We counted each answer's sources, grouped them by website (the host name without www), and recorded whether an answer cited at least one page from Reddit, Naver Blog, Tistory, YouTube, a brand in the category or a shopping site.
How many sources each engine cites
The engines differ most in how much they cite:
| Market | Engine | Answers | With sources | Sources per answer | Websites per answer |
|---|---|---|---|---|---|
| South Korea | ChatGPT | 71 | 53 (75%) | 12.6 | 6.8 |
| South Korea | Gemini | 71 | 41 (58%) | 7.3 | 3.6 |
| South Korea | Perplexity | 71 | 71 (100%) | 14.8 | 9.8 |
| South Korea | Naver AI Tab | 71 | 71 (100%) | 66.7 | 14.5 |
| United States | ChatGPT | 65 | 65 (100%) | 22.2 | 10.3 |
| United States | Gemini | 65 | 53 (82%) | 11.0 | 5.8 |
| United States | Perplexity | 65 | 65 (100%) | 13.4 | 10.2 |
Naver AI Tab lists the most links by far, up to 244 in one answer, and 58% of its links point to product pages, almost all on Naver Shopping. ChatGPT cited sources in all 65 US answers and in 53 of 71 Korean answers. Gemini answered 42 of its 136 questions with no source, and 28 of those 42 were travel and flight questions.
Same question, different websites
Put the three engines side by side on one question and their source lists rarely meet. Across the 136 questions, ChatGPT, Gemini and Perplexity cited 2,769 websites, counting each website once per question. 88% of them appeared in one engine's answer only, 10% in two and 1.8% in all three.
Pairwise, ChatGPT and Perplexity shared zero websites on 32% of the 118 questions where both cited sources, and on a typical question 8% of the websites the two cited were common to both. Gemini and Perplexity overlapped the most, at 13%. Naver AI Tab and ChatGPT shared zero websites on 70% of the Korean questions both answered with sources.
For GEO, a page one engine cites is evidence about that engine. Track citations per engine, and read one assistant's source list as a sample of that assistant.
Where each engine goes for sources
The websites each engine cited most, with the share of its answers that cited them:
| Market | Engine | First | Second | Third |
|---|---|---|---|---|
| South Korea | ChatGPT | v.daum.net 14% | card-gorilla.com 10% | card-hub.kakaopay.com 10% |
| South Korea | Gemini | youtube.com 20% | card-gorilla.com 13% | ajd.co.kr 11% |
| South Korea | Perplexity | blog.naver.com 73% | brunch.co.kr 31% | v.daum.net 17% |
| South Korea | Naver AI Tab | blog.naver.com 97% | cafe.naver.com 76% | product.shopping.naver.com 56% |
| United States | ChatGPT | nerdwallet.com 18% | rtings.com 18% | tomsguide.com 18% |
| United States | Gemini | forbes.com 18% | pcmag.com 17% | bankrate.com 14% |
| United States | Perplexity | reddit.com 34% | forbes.com 29% | nytimes.com 22% |
- ChatGPT leans on publishers and review sites. NerdWallet, RTINGS and Tom's Guide each appeared in 18% of US answers, followed by WIRED, CNBC and NBC News. In Korea its most cited site was Daum News (v.daum.net, 14%).
- Gemini cites YouTube more than any other engine: one in five of its Korean answers linked a YouTube video.
- Perplexity mixes communities and publishers: Reddit, Forbes and The New York Times in the US, and Naver Blog, Brunch and Daum News in Korea.
- Naver AI Tab stays inside Naver. Naver Blog, Naver Cafe, Naver Shopping and Naver News lead its list.
Communities, brands and shops
The share of answers that cited at least one page of each kind:
| Market | Engine | Naver Blog | Tistory | YouTube | Brand sites | Shopping | |
|---|---|---|---|---|---|---|---|
| South Korea | ChatGPT | 0% | 0% | 31% | 0% | 30% | 7% |
| South Korea | Gemini | 7% | 0% | 45% | 20% | 14% | 10% |
| South Korea | Perplexity | 15% | 73% | 42% | 7% | 49% | 17% |
| South Korea | Naver AI Tab | 0% | 97% | 17% | 0% | 24% | 56% |
| United States | ChatGPT | 0% | 0% | 0% | 0% | 46% | 23% |
| United States | Gemini | 9% | 0% | 0% | 6% | 29% | 31% |
| United States | Perplexity | 34% | 0% | 0% | 0% | 58% | 18% |
Brand sites are the websites of the brands each category tracks, such as cerave.com in skincare or kayak.com in flights. Shopping covers Naver Shopping, Amazon, Coupang, Gmarket, Olive Young, Ulta, Best Buy and other retailers.
- Reddit is a Perplexity channel. It appeared in 34% of Perplexity's US answers, 9% of Gemini's and zero ChatGPT answers.
- Naver Blog reaches Naver AI Tab (97%) and Perplexity (73%), and zero ChatGPT or Gemini answers. Tistory reaches all four engines in Korea, from 17% on Naver AI Tab to 45% on Gemini.
- Perplexity cites brand websites most (49% in Korea, 58% in the US), and Gemini least (14% and 29%).
- Shopping pages are central on Naver AI Tab, in 56% of its answers, and appear in 7-31% of answers on the other engines.
What to do with this
- Measure every engine your buyers use. One engine's citations cover a small slice of what the others cite, so a source strategy built on one assistant leaves the rest to chance.
- In Korea, publish where each engine reads. Naver Blog and Naver Cafe feed Naver AI Tab and Perplexity; Tistory, Brunch and news sites reach ChatGPT and Gemini as well.
- In the US, earn coverage on the review sites and publishers ChatGPT cites, and take part in the Reddit threads Perplexity cites.
- Keep your own product pages factual and current. Perplexity cited a brand's site in over half its answers, and those pages are the part of the source list a brand controls.
- Follow the pages one by one. A monitor reports the domains and pages each engine cites for your questions, and Answers returns every source with its position.
Method and limits
Every figure comes from one run per category with one answer per question and engine, so each engine and market rests on 65-71 answers. A 34% share from 65 answers has a 95% interval of 24-46%, and zero in 136 answers puts the 95% upper bound at 3%. The categories are buying questions in seven consumer markets, and other kinds of questions draw other sources. Websites are grouped by host name, so blog.naver.com and cafe.naver.com count as two websites. Brand sites count the brands with a known website; US supplements had few.
The answers and their sources are public. Every AI Index category opens a report with the answer behind each number, and the board's data is JSON:
curl https://api.querying.ai/v1/ai-indexFrequently asked questions
Which websites does ChatGPT cite most?
In the September 2026 AI Index run, NerdWallet, RTINGS and Tom's Guide led in the US, each in 18% of answers, and Daum News led in Korea at 14%. ChatGPT's 136 answers cited publishers, review sites, brand pages and Tistory blogs, and zero Reddit, Naver Blog, YouTube or Wikipedia pages.
Does Perplexity cite Reddit?
Yes. Reddit appeared in 34% of Perplexity's US answers and 15% of its Korean answers, more than on any other engine. Gemini cited it in 8% of answers and ChatGPT in zero.
Why does Naver AI Tab cite so many sources?
Naver AI Tab answers with Naver Blog and Cafe posts and product listings next to the text. It averaged 66.7 links per answer, and 58% of those links were product pages.
Do AI engines cite brand websites?
Yes, Perplexity most of all: 54% of its answers cited the website of at least one brand in the category. ChatGPT did in 38% of answers, Naver AI Tab in 24% and Gemini in 21%.
How can I see which pages AI answers cite for my brand?
Create a monitor with your buyer questions. Each run records every cited page per engine, and its sources report ranks domains and pages by how many answers cite them.
