Where does ChatGPT get its information?
The state of play in October 2026, briefly explained: which model answers, when ChatGPT searches the web, where the results come from and what has changed since August. With sources, and with what nobody knows for sure yet.
🇩🇪 Diesen Ratgeber auf Deutsch lesenChatGPT answers many questions from what it learned in training. For current and local questions it searches the web, using its own index and bought-in Google results. Since August 2026 it cites far less Reddit and goes straight to official websites more often.
Which model answers?
Since 7 and 8 October 2026, ChatGPT runs on GPT-6. Which version you get depends on your plan.1
| Plan | Model |
|---|---|
| Free and Go | GPT-6 Luna |
| Plus, Pro, Business, Enterprise | GPT-6 Sol |
The free plan runs on GPT-6 Luna, and Luna is by far the most used model, because about 95% of ChatGPT users don't pay.2 So what most people see in ChatGPT comes from Luna.
Two routes to an answer: memory or web search
ChatGPT has two sources. The first is what the model learned in training. It is built in, often months old and comes without sources. The second is a web search during the question. Only then are there links and sources.
ChatGPT decides which route to take itself, question by question. Mostly it does not search: according to Ahrefs, only 14.4% of free answers without login triggered a web search at all, measured on 66.6 million answers in August and September 2026.3 According to OpenAI, GPT-6 is better than its predecessors at deciding when a search is needed.4
For businesses this means: ChatGPT usually answers general questions like “What does a tax advisor do?” from memory. Questions about current things, prices or businesses in a city trigger a search. In our study, ChatGPT searched on 6 of 6 local questions.5 Small businesses get into the answer almost only through this second route.
How the web search works
ChatGPT turns one question into one or more background searches, known as “query fan-out”. According to Ahrefs, it averaged 1.6 searches per free answer that searches at all.3 Local questions can produce more: in June, ChatGPT showed us eight searches for a tax advisor in Vienna's 1st district, split by district, service and language.5
Since summer 2026, ChatGPT more often searches within a specific website using the “site:” operator, mostly the brand's own site. How often varies widely by measurement, between 6.8% and 76% of searches.36 The search words are shifting too. According to Ahrefs, “official”, “price” and “reviews” appear more and more often.3
Then ChatGPT reads the pages it found, writes the answer and adds sources. Not every page it reads gets cited, and not every cited business gets recommended.
Where the results come from
“ChatGPT searches via Bing” is no longer accurate. OpenAI itself names its own crawler and search services from other providers, including Microsoft, but does not disclose the exact mix.6 It got more specific in September. Peec AI analysed the traffic of the ChatGPT website and reports:7
- Its own index, internally called “Labrador”, actually a family of sub-indexes, for example for web, news, local businesses and shopping.
- Bought-in search results through several service providers, including Google results and Google Maps listings.
- Other partners: Microsoft's search service “Web IQ”, and Yelp and TripAdvisor for local listings.
- Ongoing tests that prefer the own index, on about 8% of chats in August.
OpenAI has not confirmed this, and Peec AI sells analytics tools itself. But we have seen two things ourselves. The settings of the ChatGPT website contain a switch called labrador_web_enabled, set to on. And the business data in ChatGPT's answers carries Google IDs and links from Google Business Profiles.5
What changed on 8 August
The biggest change did not come with GPT-6, but two months earlier, together with GPT-5.6. According to Promptwatch, it looked like this:6
| Metric | before 8 August | after |
|---|---|---|
| Share of “site:” searches | 0.37% | 16.8% |
| Searches per answer | 1.08 | 1.83 |
| Reddit's share of citations | 3.83% | 0.52% |
Ahrefs sees the same pattern over a longer period, for US answers from March to September 2026. Reddit fell from about one in five answers to one in 43, Wikipedia from one in seven to one in 35. Official sources rose; the medical database PubMed, for example, climbed from 64th to 4th among the most cited sites.3
In Vienna we measured the switch ourselves. For the question about the best restaurants, ChatGPT relied on Google reviews and Reddit in June, and on Michelin, Gault Millau and Falstaff in September.8
What GPT-6 changed
According to OpenAI, three things:4
- GPT-6 is better at deciding when it needs to search and more reliably finds information that supports the answer.
- On questions with a web search, GPT-6 Instant starts answering 44% sooner on average than GPT-5.6 Instant.
- Answers can now contain maps, tables and clickable forms. OpenAI calls this “Intelligent UI”.
There are hardly any independent measurements of GPT-6 yet. In our own tests one week after launch, GPT-6 had not reversed the August switch: no Reddit, but businesses' own websites and data from Google profiles instead.5
Which crawlers ChatGPT sends
OpenAI sends out several crawlers, each with its own job. Each can be allowed or blocked separately in robots.txt.9
| Crawler | Job | If blocked |
|---|---|---|
| OAI-SearchBot | brings websites into ChatGPT's search results | The site no longer appears in ChatGPT's search answers, at most as a navigational link. |
| ChatGPT-User | fetches a page when a user asks for it in the chat, no automatic crawling | According to OpenAI, robots.txt may not apply here, because a person triggers the fetch. |
| GPTBot | collects material for training future models | The content should not be used for training. Search is not affected. |
| OAI-AdsBot | only checks pages submitted as ads in ChatGPT | only relevant for advertisers |
So if you only want to opt out of training, block GPTBot, but not OAI-SearchBot. According to OpenAI, robots.txt changes take effect after about a day.9
What this means for businesses
- Put facts on your own website, as text: services, prices, district, languages, opening hours. ChatGPT increasingly searches official sites directly.
- Keep your Google Business Profile complete. Its data ends up in ChatGPT's answers, even though Google is not shown as the source.
- Get mentioned by sources ChatGPT trusts: industry directories, guides, press, professional chambers. Reddit is currently no lever for ChatGPT.
- Do not block OAI-SearchBot.
- Measure repeatedly, not once. The same question returns different businesses just minutes later.10
What nobody knows for sure
- Whether and how often the own “Labrador” index supplies answers. OpenAI says nothing about it.
- How big the effects really are. The figures come from tool vendors and differ widely, by more than tenfold for the “site:” share.
- How GPT-6 will behave over time. It has only been live for a week, and according to Peec AI the search system is being rebuilt week by week.7
Quick answers
Where does ChatGPT get its information?
From two sources: what the model learned in training, and a web search during the question. ChatGPT uses web search mainly for current, local or very specific questions. Only then does it show sources.
Does ChatGPT use Google or Bing?
Both, but not only. ChatGPT uses its own crawler and search services from other providers, including Microsoft. According to Peec AI, ChatGPT also buys Google results through service providers and is building its own index called Labrador. In our Vienna tests, the business data carried Google IDs.
Which model does the free version of ChatGPT use?
Since 8 October 2026, GPT-6 Luna. Paid plans get GPT-6 Sol. Because about 95% of users don't pay, Luna is by far the most used model.
Does ChatGPT search the internet for every question?
No. According to Ahrefs, only about one in seven free answers searches the web. When asked for businesses in a city, however, ChatGPT searches practically every time, in our tests on 6 of 6 questions.
Does ChatGPT still cite Reddit?
Hardly. Since a change on 8 August 2026, Reddit's share of ChatGPT citations fell from 3.8% to 0.5% according to Promptwatch (US data, provisional), and Ahrefs measures the share of answers citing Reddit at minus 88% from March to September (updated 15 Oct 2026). In our six Vienna tests in October, Reddit did not appear at all.
How does my business get into ChatGPT?
Through web search: with a website that clearly states services, prices and location in text, a complete Google Business Profile, and mentions in sources ChatGPT trusts. Do not block the OAI-SearchBot crawler. There is no guarantee, because the answers fluctuate.
Does ChatGPT find your business?
We check for free whether ChatGPT, Perplexity, Gemini and Google name you, with which details, and who gets recommended instead.
get a free AI visibility check →Sources
- Models per plan: Unite.AI, OpenAI Brings GPT-6 With Intelligent UI to All ChatGPT Tiers (October 2026).
- User numbers: TechCrunch, ChatGPT reaches 900M weekly active users (27 February 2026), with 50 million paying consumer subscribers according to OpenAI. Our calculation: 50 million out of 900 million to about 1 billion weekly users pay, so 94 to 95% don't. Counting the 9 million paying business users, it is 93 to 94%.
- Ahrefs, Louise Linehan, How ChatGPT Does Query Fan-Out (15 Million Queries Studied) (12 October 2026), citing further measurements of the “site:” share by Nectiv, Peec AI and Wil Reynolds.
- OpenAI, GPT-6 announcement of 7 October 2026, reposted in the OpenAI Developer Forum, and Search Engine Journal, ChatGPT Gets GPT-6 And Intelligent UI For Interactive Answers (7 October 2026).
- KIGEO, own measurement: How does ChatGPT find its information? Before and after GPT-6 (13 October 2026).
- Promptwatch, How ChatGPT Searches the Web (updated 25 September 2026).
- Peec AI, Tomek Rudzki, ChatGPT built its own search index (4 September 2026).
- KIGEO, AI Visibility Index, edition 2 (4 September 2026) and restaurant study (10 June 2026).
- OpenAI, Overview of OpenAI Crawlers (accessed 13 October 2026).
- KIGEO, Wie verlässlich sind KI-Empfehlungen? (How reliable are AI recommendations?, 6 August 2026, in German).