Where ChatGPT, Claude and Gemini get their answers
When an AI assistant recommends a dentist, a CRM or an app, it usually shows the pages it used. Those pages decide who gets named. So we asked three assistants the same 120 questions and sorted every site they cited.
The short version: each assistant leans on a different kind of source. ChatGPT (through Codex) mostly cited the official websites of the businesses and products it named: 63% of its cited sites. Claude (through Claude Code) and Gemini mostly cited articles and “best of” lists: 58% and 61%. For local businesses, directories and review platforms mattered most: 67% of local answers cited at least one. And the three rarely cited the same sites: 90% of the 2,048 sites in the study were cited by one assistant only.
What we asked
120 questions, on 20 November 2026, once per assistant, with web search on:
- Local services (54): a dentist, plumber, cafe, hair salon, lawyer and gym in New York, London, Sydney, Toronto, Singapore and Berlin; the same six in Gangnam, Seoul, asked in English and in Korean; and the six in Hongdae, Seoul, in Korean.
- Small-business software and services (28): for example “What’s the best AI receptionist for a small business?”, “Can you recommend a good AEO agency?” and “What tools can I use to check if ChatGPT can see my website?”. Four in Korean.
- Consumer apps (22): for example “What’s the best AI companion app?” and “Which app should I use to follow K-pop artists and sports teams for updates?”. Four in Korean.
- How-to (16): for example “How do I set up DMARC for my domain?” and “How do I fix a leaking kitchen faucet?”. Two in Korean.
Each question ended with “If you recommend or mention specific businesses, products, apps or services, put each name in bold”, so we could count names.
The assistants:
- ChatGPT through Codex with web search (model gpt-6-luna),
- Claude through Claude Code with web search (the plan’s default model),
- Gemini through Google’s API with Google Search grounding (gemini-2.5-flash).
That’s 360 answers, none failed. Every cited site is in the data file (CSV), one row per answer and site, with the kind of site we sorted it into.
Finding 1: each assistant leans on a different kind of source
Share of the sites each assistant cited (a site counted once per answer), and in brackets the share of its 120 answers that cited that kind at least once:
| Kind of site | ChatGPT | Claude | Gemini |
|---|---|---|---|
| Official website of a business or product | 63% (85%) | 7% (32%) | 18% (70%) |
| Blogs, articles and “best of” lists | 6% (20%) | 58% (93%) | 61% (91%) |
| Directories and review platforms | 10% (26%) | 18% (57%) | 10% (52%) |
| News and magazines | 4% (14%) | 12% (55%) | 5% (38%) |
| Government and professional bodies | 13% (30%) | 3% (13%) | 1% (8%) |
| Sites cited per answer | 3.6 | 7.9 | 12.2 |
ChatGPT cited few sites, and mostly the official sites of what it recommended. Claude and Gemini cited more sites, and most were articles comparing options.
Some kinds of site we expected hardly appeared:
- Reddit: 21 citations from Gemini (18% of its answers), 1 from ChatGPT, none from Claude.
- YouTube: 24 from Gemini (20% of its answers), 1 from ChatGPT, none from Claude.
- Wikipedia: 3 citations in all 360 answers.
- Social networks and map listings: under 1% of citations each.
Finding 2: what’s cited depends on what you ask
| Kind of site | Local services | Software | Apps | How-to |
|---|---|---|---|---|
| Blogs, articles and lists | 34% | 62% | 65% | 60% |
| Official websites | 25% | 19% | 16% | 23% |
| Directories and review platforms | 24% | 9% | 3% | 1% |
| News and magazines | 7% | 6% | 9% | 6% |
| App stores | 0% | 1% | 5% | 0% |
| Government and professional bodies | 6% | 0% | 0% | 5% |
Shares of all sites cited, all three assistants together (162, 84, 66 and 48 answers).
- Local services leaned on directories and review platforms: 67% of local answers cited at least one. The most cited sites were Time Out (23 answers), Fresha (10), Honeycombers (10) and, in Seoul, Daangn (9). For lawyers, legal directories such as Chambers, Legal 500 and Best Lawyers (7 answers each).
- Software answers cited comparison articles most: Fit Small Business (13 answers), TechRadar (10), Capterra, Forbes, G2 and Zapier (9 each). Some lists were published by a company selling one of the listed products, such as Toast’s own article on the best cafe POS systems.
- Apps: the App Store was cited in 16 answers and Google Play in 7.
- How-to answers cited the official help pages of the product in question, such as Google’s and Apple’s, next to articles and videos.
One detail: Claude cited dev.to, a site where anyone can publish, in 15 of its 120 answers, mostly “best X in 2026” posts. At least one was written by a company comparing itself with its competitors.
Finding 3: the assistants rarely cite the same sites
For the same question, the sites two assistants cited barely overlapped: on average 4% to 7% of the sites the two cited together were shared (Jaccard overlap, 120 questions). Of the 2,048 different sites cited in the study, 90% were cited by one assistant only, and 36 were cited by all three.
Most answers named something: 93% of ChatGPT’s answers put at least one business or product name in bold (3.4 on average), 94% of Claude’s (7.1) and all of Gemini’s (13.7).
Finding 4: in Korean, different sites, and almost no Naver
We asked about the same six kinds of business in Gangnam, Seoul, in English and in Korean. The answers cited different sites: on average 13% overlap between the English and Korean answer from the same assistant (18 pairs).
- Korean sites (.kr addresses and Korea-based services such as Daangn, Soomgo, Modoodoc, Tistory and Kakao) were 22% of the sites cited in English, and 62% in Korean (Hongdae, in Korean: 63%).
- Naver, Korea’s biggest search portal, was almost absent. None of the 36 Korean local answers cited a Naver page: no Naver Blog posts, no Naver Place listings. Across all Korean answers, Naver appeared 6 times out of 420 cited sites: five help pages for the question about registering a shop on Naver Smart Place, and one press release.
- Gemini’s Korean answers cited Tistory blogs instead.
For a business in Korea, this suggests a strong Naver presence doesn’t by itself show up in these assistants’ answers. The sample is small (18 answers per Seoul group), so treat it as a first look.
What this means for a business
- Look at what each assistant cites for your kind of question. The sources differ by assistant, by kind of question and by language. Ask your own question in each one and note the sites it shows. Our earlier study of local recommendations found the same: the directories that matter differ by country and trade.
- Keep your own website readable by AI crawlers. ChatGPT’s answers relied mostly on official sites, and Gemini cited one in 70% of its answers. Our free AI crawler check shows in a minute whether the crawlers can read yours.
- Get into the lists and directories for your trade. For Claude and Gemini, most cited pages were articles and “best of” lists; for local businesses, directories and review platforms.
- Check more than once. One answer is a snapshot: in our earlier study about half the businesses named changed when we asked again.
How we counted, and the limits
- One answer per question and assistant, on one day. AI answers change from run to run; this is a snapshot.
- Three assistants, not four. Perplexity isn’t included. ChatGPT and Claude were asked through Codex and Claude Code, which are close to, but not the same as, the ChatGPT and Claude apps.
- How sources were read: Codex and Claude Code list the pages they used at the end of the answer (we asked for that list); Gemini’s API reports the search results the answer was based on. So the number of sites per answer isn’t measured the same way. We checked a sample of 60 links Codex and Claude Code listed: 41 opened, 14 refused automated requests, 3 were not found (all from Codex), 1 didn’t respond.
- Sorting the sites: a site is counted once per answer. We sorted 76% of them with hand-written rules (known sites, government addresses, app stores, page addresses like /blog/ or “best-”, and the address matching a name the answer gave). The remaining 24% were sorted by Gemini from the address alone, without opening the page; we checked 45 of those labels by hand and agreed with 39. Where an article ends and a news site begins, or when a company’s own “best tools” article counts as a list, are judgment calls.
- Small groups: the Korean results rest on 18 answers per Seoul group and 30 answers to Korean software, app and how-to questions.
The code that asked the questions and sorted the sites is part of the OOO Works visibility engine. If you use the data, please link to this page.