Same Question, Different Sources in AI Answers

🗂️Source Types#

The differences came after general websites

General websites, including corporate, shopping, wiki and overseas sites, were the most frequently cited sources for Google AI Overviews, Gemini and ChatGPT, accounting for 55.7-78.5%. Differences between services appeared in the categories that followed. News sites accounted for just 1.4-3.5% across all four services. Sources were classified by website address alone.

Naver AI Briefing#

58.0% came from within Naver

Of the sources cited by Naver AI Briefing, 58.0% were content within Naver, including Naver Blog (46.5%) and Naver Cafe and Knowledge iN (10.4%). Naver Blog accounted for 7.7% of Google AI Overviews sources and did not appear in any Gemini or ChatGPT answers. The report concluded that operating primarily through Naver Blog was unlikely to make a site a source for those two services.

Google AI Overviews#

YouTube and social media accounted for 29.4%

Google AI Overviews drew 29.4% of its sources from YouTube and social media URLs, with YouTube making up most of that share. This was the largest category after general websites (55.7%). It also repeatedly cited the same article across multiple queries. The most frequently cited comparison article appeared as a source 411 times across 61 queries.

ChatGPT and Gemini#

Public institutions: 18.0%, blogs outside Naver: 13.1%

Public institution and organization websites accounted for 18.0% of ChatGPT's sources, the highest share among the four services. Government ministry press releases and legal information sites were among its most repeatedly cited sources. Blogs outside Naver, such as Tistory and Brunch, accounted for 13.1% of Gemini's sources. General websites accounted for 75.0% for ChatGPT and 78.5% for Gemini.

🔎Overlap with Search Results#

How closely rankings tracked citations

The share of cited sources that also appeared on the first search results page for the same query at the same moment varied widely by service. Naver AI Briefing and Google AI Overviews closely tracked search results, while ChatGPT selected sources largely independently of search rankings. Some areas were missing from the recorded first page, so these rates underestimated the actual overlap.

Two Services Close to Search#

Naver: 50.2%, Google: 40.5%

Of the sources cited by Naver AI Briefing and Google AI Overviews, 50.2% and 40.5%, respectively, appeared on the first search results page at that moment. Among articles ranked 1st on Google's first page, 32.3% were cited, compared with 13.9% of articles ranked 7th or lower. The report interpreted appearing on the first search results page as a prerequisite for citation by these two services.

Two Services Further from Search#

Gemini: 21.3%, ChatGPT: 4.2%

Gemini's sources overlapped by 21.3% with Google's first page for the same query, about half the level of Google AI Overviews. In a separate sample of 90 queries for three brands, ChatGPT's sources overlapped by only 4.2% with Bing's first page and 1.7% with Google's first page. The report suggested that ChatGPT picks its sources by running its own searches several times, and said that improving search rankings alone was unlikely to bring citations.

🔁Same Question, Different Lists#

Sources that changed or were missing

Repeating the same query did not produce an identical source list. The degree of variation differed by service, and some answers had no links at all. The report recommended measuring the same queries periodically to track trends instead of judging from a single result. This is why readers preparing a bibliography should not copy the list as it stands.

Sources Reused from the Previous Answer#

From 74.5% to 16.0%

The rate of citing the same URL as in the immediately preceding answer to the same query was 74.5% for Naver AI Briefing, 51.4% for Google AI Overviews, 36.2% for Gemini and 16.0% for ChatGPT. A URL was cited an average of 5-6 times by Naver and Google, compared with 2-3 times by Gemini and ChatGPT. The former two services repeatedly cited the same articles across multiple queries, while the latter two drew on a more dispersed set of sources.

Asked Again Three Hours Later#

Source overlap: 33.1% and 23.3%

When Gemini and ChatGPT were each asked 405 queries twice, about 3 hours apart, the citation lists overlapped by 33.1% for Gemini and 23.3% for ChatGPT. Whether an answer mentioned a brand also changed in about 1 in every 6-7 cases. The report decided to treat differences in the next round as no change if they fell within this range of variation.

Answers Without Links#

35.8% of Gemini answers

The share of answers with no source links was 35.8% for Gemini, 6.6% for ChatGPT and 1.5% for Google AI Overviews, while Naver AI Briefing had none. In Google AI Overviews, 34.9% of sources displayed only a site name without a link. The number of sources per answer was 5.9 for Google AI Overviews, 5.2 for Naver AI Briefing, 4.6 for ChatGPT and 2.4 for Gemini.

📚What Readers Should Do#

Follow links back to the original

If you copy a link list that may change when the question is repeated straight into a bibliography, readers may be unable to find the same material again. Open each link to check that the document actually contains what the answer cited, and if the page reproduces another article, trace it back to the original publisher. Find the original even when a source shows only a site name, and leave out information whose source you cannot ultimately locate.

Using AI as a Search Engine#

APA says to cite the original material you read

In an article dated September 9, 2025, the APA Style team at the American Psychological Association said that in most cases, AI need not be cited when used like a search engine to find information. It explained that this was like citing material found through Google Search: you cite what you read, not the search results. It also urged readers to open the sources suggested by AI to check that the material actually exists, and not to stop researching after asking AI.

AI Information as a Secondary Source#

South Korea's university AI ethics guidelines

South Korea's Ministry of Education "Guidelines on the Ethical Use of AI in Universities", issued on August 24, 2026, stated that checking AI-generated information against reliable sources was essential when working with academic material. Its verification examples said to treat AI-generated information as secondary material. It also said that using AI for an assignment without disclosing that use could be considered the same kind of problem as failing to cite a source.

🏷️What Site Owners Should Check#

Conditions observed alongside citations

The report measured which articles were cited and when brands were mentioned in answers to questions without brand names. It compared citations of brands' own channels, article structure and publication timing, and brand names within cited articles. These were observed associations, so they should not be read as evidence that following them will increase citations. The report also stated that it had not measured the effects of its recommendations.

When a Brand's Own Channels Were Cited#

Brand mentions: 83.7-92.3%

Among answers citing a brand's own channels, such as its official website or blog, the share mentioning that brand was 83.7% for Google AI Overviews, 90.6% for Gemini and 92.3% for ChatGPT. The mention rate in answers that did not cite the brand's own channels was 18.0-29.6%. However, answers citing those channels accounted for about 5-19% of all answers, and the mention rate for Naver AI Briefing was just 31.9% even when those channels were cited.

Content over Structure#

Subheadings, tables and FAQ offered no advantage

Across 2,297 articles on 400 first pages of Google results, cited and uncited articles from the same page were paired: the article with more subheadings was cited in 46.3% of pairs, the one with more tables in 46.9%, and the one with FAQ structured data in 46.2%. With no difference, the rate would be 50%. There was no evidence that structure offered an advantage, while Google AI Overviews tended to cite more recent articles.

To Be Cited, Content Must First Be Read#

Control search and training crawlers separately

OpenAI says that sites blocking OAI-SearchBot in robots.txt do not appear in ChatGPT search answers and may appear only as navigational links. GPTBot, which is used for training, can be allowed or blocked separately. Google's Google-Extended controls whether content is used for Gemini model training and to ground answers in Gemini apps, but does not affect inclusion in Google Search.

🔬Study Scope#

What was observed and what was not established

From August 26 to September 24, 2026, BrAInd (Jeokjeong Marketing Lab) collected 21,225 AI answers and 85,392 linked sources for 635 queries without brand names, drawn from the queries it tracked for 26 brands. This observational study examined associations and did not establish causes. Results may differ across industries or periods.