Where does ChatGPT get its information from? ChatGPT gets its information from two places: its training data, a fixed snapshot of text that ends at the model’s knowledge cutoff, and the live web, where it runs its own web searches and reads the pages they pull back. When you ask something that would benefit from current information, it rewrites your question into one or more targeted queries, sends them to search providers, and picks sources from the pages that come back.
ChatGPT may also use saved memories and earlier messages in the conversation, but that’s the smaller half. An AI visibility platform records the half that decides everything, the queries each assistant ran and the pages they pulled back, so you can see where an answer’s information actually came from. In our Credibility Layer research we logged the searches ChatGPT actually ran, and most of the pages it reads never make it into the answer.
Where ChatGPT’s information comes from
ChatGPT’s information comes from three places: its training data, the web pages its searches pull back, and what it knows about you from memory and the conversation so far. So where ChatGPT’s information comes from in a given answer is decided by whether it searched.
Training data is the base layer. It’s a fixed snapshot of text that stops at the model’s knowledge cutoff, so nothing published after that date is in it. The cutoff belongs to the model and moves with every release, and OpenAI publishes the current date on each model page in its platform docs.
The web layer turns on per question. When a question would benefit from current information, ChatGPT can search the web automatically instead of answering from training data alone, and OpenAI said it “will choose to search the web based on what you ask” when it introduced ChatGPT search. So the same product answers from memory one minute and from live pages the next.
The third place is you: saved memories and earlier messages in the conversation can shape both what it says and the searches it runs.
How ChatGPT searches the web
ChatGPT searches the web in steps: it decides to search, rewrites your question into targeted queries, sends them to search providers, reads what comes back, and may then send narrower follow-up queries.
Per OpenAI’s help article on web search, ChatGPT “may search the web automatically when your question would benefit from current information”, and it “typically rewrites your query into one or more targeted queries that it sends” to its search providers. After reviewing the first results, it “may send additional, more specific queries”. Those queries go to third-party search providers and to content supplied directly by OpenAI’s partners, so some of what looks like search results is actually partner content.
The help article’s own example is a biotech researcher asking about the latest on drugs that target CCR8 for cancer. The first query out the door is something like “CCR8 immunotherapy drug development 2025”, and after reading what returns, it fires a narrower one like “CHS-114 conference 2025”. One question became at least two searches, and the second only exists because the first came back.
Your general location, worked out from your IP, and your saved memories can shape the rewritten queries, so the same question can search differently for two people. The rewritten queries are what the industry calls fanout queries, and Omnipresence logs exactly these, in run order, joined to the pages each one pulled back.
How ChatGPT picks which sources to use
How does ChatGPT choose its sources? It picks from whatever its web searches pull back, and there’s no public formula for the picking, so the measurable part is the behaviour around it.
In the Credibility Layer study we watched it choose, across 200 reasoning sessions and 5,933 logged searches. About 30 web searches and seven sources go into a single recommendation, and about two-thirds of the searches never become visible citations, so it reads far more pages than it shows you. About 18% of its queries used site: operators to go straight to a domain.
It never ran the buyer keyword itself, in any of those searches. What it was doing was checking the business’s credentials.
And what it picks doesn’t come from Google’s top ten. In our 10x AI Visibility Refresh study, Google rank and AI citation had basically no relationship over 91 days (Spearman rho 0.000, p = 0.914), and 41 cited pages had no Google impressions at all. In the Credibility Layer data, 80% of the businesses ChatGPT recommended ranked nowhere in the Google or Bing top 50, so where it gets its information really is its own process.
What sources ChatGPT actually cited
| Source | Share of the 1,982 citations | What it means |
|---|---|---|
| The business’s own website | About 52% | The biggest source by far, and about 70% of those citations went to the homepage |
| Better Business Bureau | 261 citations, about 13% | One aggregator out-cited every paid directory combined |
| Editorial press and listicles | About 13.5% | Press coverage pays in search and in answers |
| Paid directories (Yelp, Angi, Thumbtack, HomeAdvisor, Avvo) | 8 citations | Effectively ignored |
| 0 citations | Never cited, across 200 sessions |
So which sources does ChatGPT cite most often? Credibility records did the work: the business’s own site, the Better Business Bureau and the press are where it went, and the pages you pay directories for got eight citations between them.
Those rows don’t add up to all 1,982, because the rest spread across professional associations, which the Credibility Layer study calls its biggest hidden layer, and smaller buckets. The honest limit is fragmentation: 95% of the sources appear in only one industry, so the citation mix in one category won’t look like the next.
Does ChatGPT cite its sources
Yes, when it searched: ChatGPT may include citations in a searched answer, and each one links to a page it actually read, with a Sources button that lists them. Answers built from training data carry none, because there’s nothing to link to.
OpenAI’s own caveat is blunt: “Search results and citations can be incomplete, outdated, or incorrect.” A citation is evidence that the model read the page, not an endorsement of what the page says, and OpenAI tells users to open the sources and check them. So being cited is the win, and being verifiably right on the page is what keeps the citation worth having.
What this means for getting your pages cited
So how do you get cited by ChatGPT? Here’s what to do about it, in the order it pays off.
First, your own site gets you read, and third-party records get you recommended. In the Credibility Layer data, no recommendation came from the business’s site alone: every one paired the site with outside checks. So the unit of work is two things together, a citation-grade page on your site plus the third-party records that back it up.
Second, write for the queries it actually runs. The searches in the study were credential checks, site: lookups on specific domains and the like, and the pages that win them answer those checks directly. Your keyword list answers different questions.
Third, make those pages answer the checks in liftable passages. If the query is a site: lookup on your domain, the page it lands on has to carry the credential in a form the model can read and quote, and the credential has to be visible without reading the whole page first.
You can spot-check a few prompts by hand and read the citations yourself. Past that, Omnipresence shows you which buyer questions ChatGPT, Claude, Gemini and Perplexity recommend you on, the searches behind those answers, and the pages that won them, and your agent gets the playbook to work the ones you’re losing. It covers those four assistants only, and the top-choice data is partial, so read it as a work list rather than a census.
FAQ
Does ChatGPT use the information you give it?
In the conversation, yes, that’s the point of it. With memory enabled, ChatGPT may carry saved details into later answers and into the search queries it rewrites. From an operator’s side, the practical read is that what’s on the page matters more than what’s in the chat: searches pull pages, and pages are the thing you can work.
Why chatgpt.com shows up in your analytics
Because someone clicked a citation link in a ChatGPT answer and landed on your site, and the visit arrives with chatgpt.com as the referrer. It’s your citations turning into visits, not ChatGPT crawling your pages. If the number shows up, treat it as a rough count of answer clicks rather than a ranking.