ChatGPT gets its information from two places. The first is its training data: a huge body of public web pages, licensed data and human feedback, collected before a cut-off date and compressed into the model. The second is live web search: when a question needs something current, ChatGPT searches the web, reads the results and cites them. A third, smaller source is you: your memories, files and connected apps. Which one answers a question decides how fresh the answer is, whether it has links, and, if you run a business, whether your brand can appear in it. This post sets out what OpenAI says about each, what it does not say, and what the independent evidence adds.
The short answer: memory or search
Think of ChatGPT as a well-read assistant with a phone. It can answer from what it has read, which is broad but dated. Or it can look something up, which is current but depends on what the search finds. Most of what you notice about an answer follows from which of the two it used.
| What it is | How current | Links to sources? | How a brand gets in | |
|---|---|---|---|---|
| Training data | Public web, licensed partner data, user and trainer data | Up to the model's knowledge cutoff | No: the model does not keep copies of pages | Being described widely and consistently before the cutoff |
| Live web search | Results from third-party search providers and partner content | Current | Yes, usually | Being crawlable, indexed and ranking for the searches it runs |
| Your own context | Saved memories, past chats, custom instructions, files, connected apps | As current as your data | Sometimes | Only through what the user has shared |
Source 1: training data
OpenAI says its models, “including the models that power ChatGPT”, are developed from “three primary sources of information: (1) information that is publicly available on the internet, (2) information that we partner with third parties to access, and (3) information that our users, human trainers, and researchers provide or generate.”[1]
The public part means content that is “freely and openly accessible”: “publicly available webpages, public forums, public blogs, public posts, and other publicly available online content.” OpenAI says it does not intentionally gather data “from sources known to be behind paywalls or from the dark web”, and filters out “hate speech, adult content, sites that aggregate personal information, and spam”. It also uses “synthetic data in some training processes”[1].
Two details matter more than the list. First, the model is not a library. OpenAI says the models “do not store or retain copies of the data they are trained on”[1]. They learn patterns, which is why an answer from memory has no links: there is no page to point to. It is also how a model comes to know a brand: OpenAI says models may learn “to recognize public figures and well-known entities”[1]. A brand the web discusses often and consistently is easier to learn than one mentioned twice.
Second, the cutoff moves with each new model, and so do the answers. Some guides still say ChatGPT’s dataset “only went up to 2021”[18]. That was true of early models and has not been for years. If you need to know what a specific model knows, read its page in OpenAI’s model documentation rather than a blog table.
Site owners have one control here: OpenAI says they can use robots.txt “to disallow GPTBot, which may crawl publicly available content to help train our models”[1]. That keeps future crawls out of training. It does not remove you from search, which is the next section. Our post on whether to block GPTBot covers the trade.
Source 2: live web search
OpenAI says “ChatGPT may search the web automatically when your question would benefit from current information.” You can also ask it to. When it searches, “ChatGPT search typically rewrites your query into one or more targeted queries that it sends those providers”, and “responses that use web search may include citations”[3]. The rewritten searches can carry context you did not type: “if memory is enabled, ChatGPT may use relevant saved memories when rewriting a search query”. OpenAI’s own example: someone who has said they are vegan and live in San Francisco asks for nearby restaurants, and ChatGPT may search for “good vegan restaurants San Francisco”. It may also share “general location” from your IP address with “third-party search providers”[3].
Searching is not the default for every question. In a 2026 replication by AI+Automation, a vendor, “only 35% of queries (136 out of 400) triggered any web search at all”[15]. That figure depends on the questions they chose, but the point holds: many answers, including many about products, come from training alone. OpenAI is also frank about search quality: “Search results and citations can be incomplete, outdated, or incorrect.”[3]
Search answers also draw on partners. At launch, OpenAI said ChatGPT search “leverages third-party search providers, as well as content provided directly by our partners”, and that it had partnered “with news and data providers” for weather, stocks, sports, news and maps[4].
Does ChatGPT use Bing?
Here is exactly what OpenAI says today. Its search help page says “ChatGPT search sometimes partners with other search providers”, and for how those providers handle queries it links two privacy policies: Microsoft’s and Shopify’s[3]. Its privacy policy refers to “search and shopping providers” without naming them[7]. The word “Bing” does not appear on either page.
It did once. When OpenAI launched browsing as a plugin in March 2023, it wrote: “Browsing retrieves content from the web using the Bing search API.”[6] That describes a feature OpenAI has since replaced, so it is history, not a statement about today’s ChatGPT search.
Outside studies try to infer the provider by comparing ChatGPT’s citations with each engine’s results. They agree on the direction and disagree on almost everything else:
| Bing overlap | Google overlap | Method | |
|---|---|---|---|
| Seer Interactive, Feb 2025 | 87% matched Bing's top 20 | 56% matched Google | 500+ citations for about 100 questions |
| Ahrefs, Aug 2025 | About 10% in Bing's top 10 | About 12% in Google's top 10 | 15,000 long-tail queries; ChatGPT, Gemini, Copilot and Perplexity counted together |
| AI+Automation, Apr 2026 | 27.4% | 13.4% | 400 questions, counting the searches ChatGPT ran |
Sources: Seer Interactive[13], Ahrefs[14], AI+Automation[15].
The honest reading: Bing looks like a main provider, OpenAI does not confirm it in its current documentation, and a large share of citations match neither engine’s results for the question as typed, partly because ChatGPT rewrites the question first. Being indexed in Bing costs little and removes one way to be missing. The Bing Webmaster Tools guide covers setup.
Does ChatGPT use Google?
OpenAI has not said so. In August 2025, The Information reported that OpenAI had been “partially using Google search results scraped by a startup called SerpApi for ChatGPT responses on current events like news and sports”, as summarised by Search Engine Roundtable[11]. Tom’s Guide reported the same story, describing data that “reportedly helps ChatGPT answer real-time queries about news, sports, and financial markets”[12]. We found no statement from OpenAI confirming or denying it. The only mention of Google on OpenAI’s search help page is a shortcut in its Chrome extension that sends a search to Google instead of ChatGPT[3].
One older academic study is often cited here. Before ChatGPT could search, researchers compared GPT-3.5 and GPT-4 answers to health questions with Google’s top results and found “the ChatGPT response was dominated by the top Google Search result”[19]. That was a comparison of text, not evidence of a live connection. The likelier explanation is that pages which rank well on Google are also widely copied and quoted, so models learn from them too.
For a brand, the practical answer is the same either way. The pages that rank well on Google and Bing for a question are the pages most likely to be found when ChatGPT searches it. Ordinary search visibility still matters.
OpenAI’s own crawlers
Whatever the search provider, OpenAI runs its own crawlers, and they do different jobs. Its documentation says “each setting is independent of the others”[8]:
- OAI-SearchBot “is used to surface websites in search results in ChatGPT’s search features. Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links.”
- GPTBot crawls “content that may be used in training our generative AI foundation models”. Disallowing it indicates your content should not be used for training.
- ChatGPT-User visits a page when a user’s question leads ChatGPT to it. It “is not used for crawling the web in an automatic fashion”, and “is not used to determine whether content may appear in Search.”
OpenAI’s help page is explicit about eligibility: “To make a website eligible for inclusion, allow OAI-Searchbot to crawl the site” and make sure your host or CDN allows OpenAI’s search IP addresses. And it says “placement is not guaranteed”[3]. Changes to robots.txt take about a day: OpenAI says “it can take ~24 hours” for search to adjust[8]. Back in 2024, OpenAI drew the line clearly: “Sites can be surfaced in search results even if they opt out of generative AI training.”[5] The free AI crawler checker shows which of these bots your live robots.txt allows.
Other sources: shopping data and your own context
Products. For shopping questions, OpenAI says ChatGPT considers “structured metadata from first-party and third-party providers (e.g., price, product description) and other third-party content”. Prices come “from third-party providers”, review summaries “are based on reviews from public websites”, and for Shopify merchants, “product data is already integrated into ChatGPT through Shopify Catalog”. OpenAI adds that product results “are not ads, nor influenced by any OpenAI partnerships”[9]. The ChatGPT Shopping guide covers feeds and merchant setup.
You. With memory on, ChatGPT “can remember relevant preferences and details from your chats and other available sources”, including saved memories, past chats, custom instructions, files and connected apps[10]. This is why two people can get different answers to the same question, which our post on why ChatGPT gives different answers covers.
What sources does ChatGPT cite?
OpenAI says only that ChatGPT “ranks search results using multiple factors intended to help users find relevant, reliable information”[3]. It publishes no list. What studies show is that the cited sources change fast and lean on third-party pages. SISTRIX tracked 82,619 prompts for 17 weeks and found ChatGPT replaces “as much as 74%” of its sources every week[16]. Forums feature heavily: Semrush found Reddit links in 12.6% of the ChatGPT search answers it sampled[17]. Our post on how AI assistants choose what to cite separates what is known about selection from what is guessed.
What this means for your brand
Because answers come from two places, there are two routes into them, and they work on different clocks.
| What helps | How fast | How to check | |
|---|---|---|---|
| Training data | Clear, consistent descriptions of you on the sites that cover your category: reviews, comparisons, forums, your own pages | Slow: shows up when a new model is trained | Ask a model without web search, many times |
| Live search | OAI-SearchBot allowed, indexed in Bing and Google, pages that answer buyer questions directly, a place on the comparison pages that rank | Fast: days to weeks | Ask in ChatGPT with search, and read the sources it cites |
The training route is the one most brands neglect, because it cannot be fixed with one page. If the web described you as “a tool for freelancers” for years, a model trained on that will keep saying so after you move up-market, until new descriptions outnumber the old and a new model is trained. The search route is quicker but rotates: being cited this week is no promise for next week.
You can check the two separately. The free ChatGPT rank tracker asks OpenAI’s and Google’s models through their APIs without web search, so it shows roughly what the model says from training. Asking the same question in the ChatGPT app, where it may search, shows the other side. Ask each a few times: one answer is a sample, not a verdict.
Tracking whether it names you
Knowing where answers come from tells you what to work on. Knowing whether it is working needs repeated measurement. CoreCited runs your buyers’ questions every week across ChatGPT and the other major assistants and records who each answer names and which pages it cites; see AI visibility tracking and citation analysis.
Questions people ask
Where does ChatGPT get its information?
From two places. First, its training data: public web pages, data OpenAI licenses from partners, and data from users, trainers and researchers, all collected before the model's knowledge cutoff. Second, live web search: when a question needs current information, ChatGPT can search the web through third-party search providers and its own crawler, and cite the pages it used.
Does ChatGPT use Bing?
OpenAI's current help pages do not name Bing. They say ChatGPT search uses third-party search providers and link Microsoft's privacy statement for how those providers process queries. In 2023, OpenAI said browsing used the Bing search API. Vendor studies find ChatGPT's citations overlap more with Bing's results than Google's, though their numbers vary widely.
Does ChatGPT use Google?
OpenAI has not said so. In August 2025, The Information reported that OpenAI had been partly using Google results scraped by a company called SerpApi for some current-events answers, as covered by Search Engine Roundtable and Tom's Guide. We found no OpenAI statement confirming or denying it.
How does ChatGPT search work?
When ChatGPT decides a question needs current information, it rewrites the question into one or more search queries, sends them to its search providers, reads the results, and writes an answer with citations. It may use saved memories and your approximate location when rewriting the query. Sites must allow OpenAI's OAI-SearchBot crawler to be eligible.
Does ChatGPT have access to the internet?
Yes, when it searches. Without search, it answers from what it learned in training, which stops at the model's knowledge cutoff. With search, it can read current pages and cite them. ChatGPT decides on its own when to search, and you can also ask it to.
How do I get ChatGPT to mention my brand?
Work on both sources. For training data, be described clearly and consistently on the sites that write about your category: reviews, comparisons, forums and your own pages. For search, let OAI-SearchBot crawl your site, make sure Bing and Google index it, and publish pages that answer buyers' questions directly. Then check repeatedly, because answers vary.