Where Does ChatGPT Get Its Information About Your Business?

Key points

• Training data is set before a model ships. OpenAI says its GPT-5 models were trained on public internet data, data from partners, and data from users, trainers and researchers.
• OpenAI runs a training crawler (GPTBot) and a separate search crawler (OAI-SearchBot). Each answers to its own robots.txt rule, so blocking training doesn't have to mean leaving search.
• In Vector Index AI's API tests from August and September 2026, ChatGPT averaged 6.1 searches per business-recommendation question, and 157 of its 478 searches checked a single site.

Where does ChatGPT get its information, source by source?

Source	What goes in	When ChatGPT uses it	What you control
Training data	Public internet data, partner data, and data from users, trainers and researchers	Fixed before the model is released	The GPTBot rule in your robots.txt (training only)
Live web search	Pages found while answering	When the model decides the question needs it	The OAI-SearchBot rule in your robots.txt
Pages a user sends it to	A page ChatGPT opens because someone asked	During a chat	Limited: OpenAI says robots.txt "may not apply"
What the user brings	Pasted text and uploaded files	During a chat	None

Training can't include anything you published after the model was built. A new service, a move or a new credential reaches ChatGPT through search or through a user who supplies it, which makes the second row the one to act on first.

What did ChatGPT learn in training?

OpenAI's GPT-5 system card says the GPT-5 models were trained on "information that is publicly available on the internet, information that we partner with third parties to access, and information that our users or human trainers and researchers provide or generate."

Partner data can be paid for. Fortune reported in December 2023 that OpenAI would compensate Axel Springer for using its content to train its models.

You can't write yourself into training data, but you can decide whether OpenAI may use your site for it. According to OpenAI's crawler documentation, disallowing GPTBot "indicates a site's content should not be used in training generative AI foundation models." That rule covers training, not search answers.

When does ChatGPT search the web for current information?

When the model judges that the question calls for it. OpenAI's web search guide for developers says the model "can choose to search the web or not based on the content of the input prompt." That page documents the API, so treat it as how OpenAI's models behave there, not a promise about every ChatGPT setting.

The same guide says "the number of sources is often greater than the number of citations." The links under an answer often aren't a full list of what ChatGPT consulted.

OpenAI documents two user agents that matter here:

• OAI-SearchBot is used "to surface websites in search results in ChatGPT's search features." OpenAI says sites opted out of it will not be shown in ChatGPT search answers, though they can still appear as navigational links. It can take about 24 hours after a robots.txt change for OpenAI's systems to adjust. OpenAI recommends allowing OAI-SearchBot in robots.txt and allowing requests from its published IP ranges.
• ChatGPT-User visits a page when someone's question needs it. OpenAI says it isn't used to decide what may appear in search, and because a user starts the visit, "robots.txt rules may not apply."

Hypothetical example: a dental practice wants to stay out of AI training, so its developer blocks every crawler in the "" section of robots.txt, with no section for OAI-SearchBot. Under the robots.txt standard, RFC 9309, a crawler with no matching section must follow the "" section. The practice has opted out of ChatGPT search answers without meaning to. A separate section allowing OAI-SearchBot, with GPTBot still blocked, gives it what it wanted.

What does ChatGPT search for when asked to recommend a business?

In its study of what AI searches for when asked to recommend a business (linked at the end), Vector Index AI recorded the searches the engines ran for themselves. ChatGPT was called as gpt-5.4 through its API with web search on, on 24 and 25 August and 16 September 2026.

• ChatGPT ran 478 searches of its own across 78 answers. That's 6.1 per question, with 12 in its busiest single answer.
• 157 of those 478 searches used the site: operator to check one domain. They targeted 97 distinct domains across 38 answers, and bbb.org was the most-targeted, with 13 searches.
• ChatGPT went from the category to named companies, then checked them one at a time.

Separately, in answers to local questions that named Cincinnati or said "near," the local AI citations study found ChatGPT cited bbb.org in 14 of its 68 answers, more than any other single domain. Only linked domains counted, and these are readings on the days sampled, not rates.

Read both with their limits. API answers differ from the consumer app in system prompt, retrieval and routing. Most questions named Cincinnati. Vector Index AI wrote the questions, so they aren't a sample of what buyers ask. And the study recorded where ChatGPT searched, not whether what it found changed the answer.

In those runs, ChatGPT's site checks went to third-party sites such as bbb.org and to companies' own domains, so what your site and profiles say is part of what its search can turn up about you. What ChatGPT looks for on your website covers the words those checks asked for.

Are the sources ChatGPT cites always right?

Not always, so open them. In a test published in Columbia Journalism Review in March 2025, the Tow Center for Digital Journalism gave eight AI search tools news article excerpts and asked each for the article's headline, publisher, date and URL. ChatGPT incorrectly identified 134 articles and signaled a lack of confidence just fifteen times out of its two hundred responses. That test covered news, and the product has changed since. A cited link still only tells you where ChatGPT says it looked, so check that the page says what the answer claims about you.

How can you check where ChatGPT gets its information about your business?

You end with a dated record of which sources ChatGPT cites for your service and town, and a robots.txt that lets it search your site.

1. Open yourdomain.com/robots.txt. If a section naming OAI-SearchBot blocks it, or there's no such section and the "*" section blocks everything, you're opted out of ChatGPT search answers. Allow it and give OpenAI about 24 hours to adjust.
2. Decide on GPTBot separately. Blocking it is a training choice and doesn't need to touch OAI-SearchBot.
3. If your host or firewall filters bots, ask them to allow requests from OpenAI's published IP ranges, so the firewall doesn't block what robots.txt allows.
4. In ChatGPT, use search and ask what a buyer would ask, such as your service plus your town. Record the exact wording, the date, the model or mode shown, every cited URL and every business named.
5. Open each cited link. Mark it as your page, a directory or review profile, or another business, and confirm it says what the answer attributes to it.
6. Ask the same question on different days. One answer is one reading.
7. Fix what you own. Correct profiles you've claimed, and make sure your site says in plain text what you do, where, and which credentials you hold.

When a buyer asks ChatGPT who to hire, part of where ChatGPT gets its information is the searches it writes for itself, which in Vector Index AI's tests were aimed at named businesses and the sites that list them. Before you change a page or claim another profile, read the searches ChatGPT wrote for itself, word for word. You'll see the exact queries it ran on recommendation questions, so you know which kinds of sites to check first.