How ChatGPT chooses its sources, and why its answers change

ChatGPT builds its answers from two places: what its model learned in training, which comes with no links, and a live web search, whose sources it cites in the reply. OpenAI's help center says ChatGPT searches on its own when a question would benefit from current information, rewrites the question into one or more targeted queries, and can adjust those queries with your approximate location and your saved memories. So ChatGPT's sources, and the brands it names, can change from one person and one session to the next.

Americans search “does chatgpt give the same answer to everyone” about 260 times a month (Google Ads estimate, October 2026). Google works from another base entirely. Google AI Mode runs its searches on Google's own index, whereas ChatGPT relies on outside search providers, partner content and its own crawler, so a brand that cares about AI answers has to watch both.

Since those sources move, a brand has to read them over time. Hypervisible, the software we publish, logs the sites ChatGPT cites on the questions buyers in your market ask, next to the citations of up to four other engines (Gemini, Perplexity and Google's AI Overviews and AI Mode), week after week.

When ChatGPT searches the web

ChatGPT searches the web when it judges that your question needs current information, and you can make it search when it doesn't. Everything else comes from what the model already knows.

OpenAI's help page on web search, read on October 22, 2026, says ChatGPT “may search the web automatically” when a question would benefit from current information. To force a search, open the tools menu and pick Search, or type a slash in the message box and choose Search. On an answer that came back without links, the regenerate control offers Search the web. Search works on every plan from Free to Enterprise, and for people who aren't signed in, within each plan's usage limits. It arrived in October 2024 for paying users and opened to everyone, with no sign-up, in February 2025.

OpenAI doesn't publish the rule that decides when a search happens. Questions about prices, recent news or local businesses need current information by definition, so they are the obvious candidates.

Without a search, the answer rests on training data: a page you published last month can't play a part until a later model learns from it, and the answer carries no link you could check. Once a search runs, a page OpenAI's crawler has read can be cited as soon as it's found.

The search itself can involve several queries. When ChatGPT works with a search provider, OpenAI says it “typically rewrites your query into one or more targeted queries.” The help page gives an example from medicine: asked about drugs that target CCR8 in cancer, ChatGPT might first search “CCR8 immunotherapy drug development 2025,” read the results, then send a narrower query, “CHS-114 conference 2025,” built from what it found. Google calls the same idea query fan-out.

Take a buyer instead (the example is ours). A practice manager in Austin asks which scheduling software suits a dental office with two locations. ChatGPT might run one search on multi-location features and another for reviews written by dentists. Each query brings back its own pages, and the answer cites the ones that settle a piece of the question.

Where the sources come from

A ChatGPT answer can draw on five things: the model's training, web results from search providers, content from OpenAI's partners, product data, and what ChatGPT knows about the person asking. Only some of them come with links.

Based on OpenAI's help center, developer documentation and launch post, read on October 22, 2026.
Where it comes from	Shown with a link	What a site or brand can do
The model's training: public web content, data from partners, input from users and trainers	No	Allow or block GPTBot, which collects content for future training
Web search through third-party search providers	Yes, as citations and under Sources	Allow OAI-SearchBot and stay well indexed on the web
Content supplied directly by OpenAI's partners	Yes	Partnership agreements, mostly with news publishers
Product data for shopping questions	Yes, on product cards with links to merchants	Keep product data accurate; Shopify stores are included, other merchants can apply for a direct feed
The user's own context: memories, custom instructions, past chats	Under Sources, for personal context	Nothing on the brand side

Start with training. OpenAI says its models learn from three main sources: information publicly available on the internet, information it accesses through third-party partners, and information that users, human trainers and researchers provide or generate. The models don't keep copies of what they read; training adjusts their parameters to reflect patterns. That's why ChatGPT can describe your company without linking to anything. No page sits behind the sentence. Site owners decide on training through GPTBot: disallowing it in robots.txt tells OpenAI the content shouldn't be used to train its foundation models. It has no effect on search.

Search works differently. In its October 2024 launch post, OpenAI said ChatGPT search relies on third-party search providers and on content provided directly by its partners. The current help page names no engine, but it sends readers to Microsoft's and Shopify's privacy policies to see how search providers handle queries. Microsoft had announced in May 2023 that Bing was coming to ChatGPT as its default search experience. The partners named at launch were mostly publishers, among them the Associated Press, Axel Springer, the Financial Times, Hearst, News Corp, Reuters and Vox Media, plus providers of weather, stock, sports and map data.

OpenAI also runs a crawler of its own for search, OAI-SearchBot. According to its developer documentation, a site that opts out of OAI-SearchBot isn't shown in ChatGPT search answers, apart from navigational links, and a robots.txt change can take about 24 hours to register. A separate agent, ChatGPT-User, opens a page when a user's request calls for it. OpenAI says robots.txt rules may not apply to those visits, and that ChatGPT-User plays no part in deciding what appears in search.

Shopping questions follow their own path. OpenAI says product results are chosen independently and aren't ads. To pick them, ChatGPT reads structured metadata such as price and description from first-party and third-party providers, other third-party content, and what the model generated before it searched. Shopify merchants are already included through Shopify Catalog; other merchants can apply to send OpenAI a direct product feed. Review summaries are written from reviews on public websites, which OpenAI doesn't verify.

Why the same question gets different answers

ChatGPT picks its words with some randomness, the searches it runs depend on who is asking and from where, and both the web and the models keep changing. Any of the three can move a brand in or out of an answer.

On randomness, OpenAI is explicit. Its article on how the models are developed explains that a sentence can be completed in several plausible ways, so there is an inherent element of randomness in each response, and the same question may yield different answers across queries. Ask twice in a row and the shortlist can shift.

Then comes context. The searches aren't fixed: ChatGPT rewrites them from your question, and OpenAI lists what can shape that rewrite.

Location. ChatGPT estimates a general location from your IP address and may pass it to search providers. Someone in San Francisco asking for good restaurants nearby may get a search for top restaurants in San Francisco. Sharing a precise device location is optional and off by default.
Memory. With Memory on, ChatGPT may use saved memories when it rewrites a search. In OpenAI's example, a user who said they are vegan and live in San Francisco may get a search for good vegan restaurants there. Depending on plan and region, Memory can draw on past chats, saved memories, custom instructions, files in the Library and connected apps such as Gmail.
Search or no search. One session searches the web, another answers from training. The first cites pages and the second doesn't.
Shopping preferences. Product picks weigh the query and context such as Memory or custom instructions. In OpenAI's example, someone who has said they dislike clowns may not be shown clown costumes.

Time does the rest. Search results move as pages are published and updated. The models change too: OpenAI describes ongoing evaluation and improvement after deployment, and a new model can reorder a list of recommendations.

For a brand, a single screenshot is the wrong unit of measure. Your presence in ChatGPT is better read as a percentage, worked out over many questions and repeated readings. Nobody can sell you a permanent spot in those answers, any more than anyone could on Google's first page, but the rate itself can be pushed up and tracked.

What makes a page easy to cite

OpenAI says only that ChatGPT ranks search results on several factors meant to help users find relevant, reliable information, and that it promises no placement. The rest is public practice, much of it the same as for search engines.

Let OAI-SearchBot in

The help page states the one explicit requirement: allow OAI-SearchBot to crawl the site, and make sure your host or CDN accepts traffic from OpenAI's published IP addresses. Check both. A firewall rule that blocks unfamiliar bots shuts ChatGPT out as surely as a line in robots.txt, and marketing may never hear about it.

Answer the narrow question

ChatGPT turns a question into narrower queries, so the page that gets picked is often the one that answers a narrow query cleanly. A section titled “Pricing for multi-location dental practices,” with the answer in its first two sentences, gives a model something to quote. A page that circles the topic for 800 words before saying anything gives it much less.

Give facts a model can check

Prices, dates, specifications and named sources give a model something firm to repeat. Research points the same way. In the 2023 paper that coined generative engine optimization, a team from Princeton University and IIT Delhi found that adding statistics, quotations and cited sources to a page raised its visibility in AI answers by up to 40% on their measures, while keyword stuffing did little.

Be named where ChatGPT looks

When someone asks ChatGPT about a category, the answer often leans on review sites, comparison articles, forums and trade press that you have no say over. For shopping questions, OpenAI says review summaries come from public websites. Getting listed and reviewed there, and described correctly, is part of the job. The citations ChatGPT gives on your own questions show which of those sites matter for you. NEXUS GEO, the agency behind Hypervisible, sets out a fuller B2B plan in how to appear in ChatGPT, and the search feature itself is covered in our guide to ChatGPT search.

How to see which sources it uses for your topic

Open the citations and the Sources panel on answers that used search, then ask the same questions again on a schedule and keep a list of the domains that come back.

On an answer that searched the web, citations sit next to the claims they support. Click one to open the page, or hover over it on desktop for a preview. The Sources button, when it appears, lists the cited pages and other relevant links. OpenAI warns that search results and citations can be incomplete, outdated or incorrect, so open the page before you trust the claim.

For your own market, turn that into a routine.

Write down the questions your buyers ask before they choose, in their words. Keep the ones that don't name you apart from those that do.
Ask them in a temporary chat and choose Unpersonalized before you start; memories and custom instructions then stay out of it. For anything local, put the city in the question, since ChatGPT otherwise guesses it from your IP address.
Check that a search ran. If the answer has no citations, regenerate it with Search the web.
Record the brands named, their order, and every cited domain.
Repeat on the same day each week. A single reading tells you little; the domains that show up week after week are where the work is.

Your server logs fill in the other side. OAI-SearchBot visits show OpenAI's search crawler reading your pages; ChatGPT-User visits show ChatGPT opening a page because someone asked.

Tracking ChatGPT's sources with Hypervisible

Kept up by hand, that routine can eat an afternoon a week, and ChatGPT is only one engine. We recommend Hypervisible for the job, and we publish it: NEXUS GEO, a French agency that specializes in GEO, built it on data from Google and the AI engines.

The source list is what fits this topic best. For every answer, Hypervisible records the sites the engine cited, so the domains that keep coming back stand out without a spreadsheet. Beside that list you get your citation rate (how often answers name you), your average position, where you rank among the competitors that share those answers, and the content and tone of what the engines say about you.

ChatGPT adjusts its searches to where you are, so Hypervisible measures market by market, in ten countries: English-language questions for the United States, the United Kingdom, Canada and Singapore, and each of the others, France and Germany among them, in its own language. ChatGPT is read on every plan, alongside Gemini and Google AI Overviews, and the five-engine plans add Perplexity and Google AI Mode. Paid plans go through the list again each Monday and show the gap since the last reading in points (pricing).

It reads once a week, so a change that lasts a day can slip through, and Claude and Copilot are outside its scope. For a trial run, the free AI visibility checker asks three engines (ChatGPT, Gemini and Google AI Overviews) 10 questions each: five on your sector that leave your name out, five that include it. Expect 5 to 10 minutes for the 30 answers; it costs nothing, asks for no card, and saves the results to your account; the scan page is another way in. How the other engines behave is covered in ChatGPT vs Gemini vs Perplexity, and more broadly in AI search engines compared.

Your brand

The free analysis asks 10 questions to three AI engines (ChatGPT, Gemini and Google AI Overviews), reads the 30 answers in 5 to 10 minutes and keeps your dashboard in your account.

Frequently asked questions

Does ChatGPT give the same answer to everyone?

No. OpenAI says its models respond with an inherent element of randomness, so the same question may get different answers. Context changes the answer too: ChatGPT may use your approximate location and, with Memory on, your saved memories when it rewrites a web search, and one session may search the web while another answers from training. For a fair comparison, use a temporary chat set to Unpersonalized and ask several times.

Where does ChatGPT get its information?

From two main places. Its models learned from information publicly available online, data from third-party partners, and input from users, trainers and researchers. When a question needs current information, ChatGPT also searches the web through third-party search providers and partner content, and cites the pages it used. With Memory on, it can add your own context, such as past chats and custom instructions.

Can you see ChatGPT's sources?

Yes, when the answer used a web search. Citations appear next to the statements they support, and a Sources button, when available, lists the cited pages and other relevant links. Answers drawn only from training have no links to show. If you want sources, regenerate the answer and choose Search the web, then open the pages to check that they support what ChatGPT wrote.

What search engine does ChatGPT use?

OpenAI doesn't name a single one. Its help center says ChatGPT search sometimes works with other search providers and points to Microsoft's and Shopify's privacy policies for how they process queries. Microsoft announced in May 2023 that Bing would be ChatGPT's default search experience. OpenAI also runs its own search crawler, OAI-SearchBot, and adds content supplied by partners such as news publishers.

Does blocking GPTBot keep a site out of ChatGPT?

Not out of its search answers. GPTBot collects content that may be used to train OpenAI's models, and disallowing it only opts your pages out of that training. ChatGPT search depends on a different crawler, OAI-SearchBot. Block that one and OpenAI says your site won't be shown in ChatGPT search answers, apart from navigational links. The two settings are independent of each other.

How can you track which sites ChatGPT cites for your brand?

Hypervisible, the software we publish, does it for you: every Monday it asks ChatGPT, and up to four other engines, the same buyer questions and records which sites each answer cites, whether your brand is named, and in what position. We recommend it because it measures by market, the US included, with questions written in English. By hand, open the citations and the Sources panel on answers that used search, and log the domains each week.