Where Does ChatGPT Get Its Information? (2026)

How often does it search? We tracked 146 B2B SaaS buying questions in ChatGPT every day for 28 days (23 September to 20 October 2026). ChatGPT ran a web search in 87.7% of its 4,088 answers, and in the last 14 of those days a searched answer read a median of 22 pages and cited a median of 3.

Training data: what OpenAI says ChatGPT learned from

OpenAI's help page on how its models are developed names three sources: "(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."

The public part is broad. It "may include 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 that it "increasingly" uses synthetic data in some training.

The model is not a library of its sources. OpenAI says the models "do not store or retain copies of the data they are trained on," and that ChatGPT does not copy and paste from training data. OpenAI also publishes no dataset list for the current models. The October 2026 system card for GPT-6 Sol and GPT-6 Luna says only that they "were trained on diverse datasets and filtered through our data processing pipeline."

Human feedback shaped how ChatGPT answers. At launch in November 2022, OpenAI said it trained ChatGPT using Reinforcement Learning from Human Feedback (RLHF): trainers wrote example dialogues, then ranked alternative answers to build a reward model.

Licensed content splits two ways. A few deals cover training: the Associated Press licensed part of its text archive, Axel Springer's deal covers training and displayed summaries, and the Financial Times deal covers display plus model improvement. Most of the rest are display deals, where ChatGPT shows summaries, quotes and links. Le Monde, Prisa Media, News Corp (including The Wall Street Journal), Condé Nast, the Guardian and The Washington Post sit in that group. In April 2025, OpenAI reposted The Washington Post's announcement of their deal, which says OpenAI has deals "with more than 20 news publishers."

What is ChatGPT's knowledge cutoff?

OpenAI's first GPT-6 models reached paid ChatGPT customers in September 2026. On 7 October 2026, the Chat tab began moving to new October versions of GPT-6, which replace GPT-5.6 Sol and GPT-5.6 Luna. The switch reached Plus, Pro, Business and Enterprise from 7 October, and Free and Go from 8 October. OpenAI's release notes say Plus, Pro, Business and Enterprise use GPT-6 Sol, and Free and Go use GPT-6 Luna.

OpenAI's system card for the October versions of these models in ChatGPT names no knowledge cutoff. OpenAI's API model pages do: Apr 20, 2026 for gpt-6-sol and May 18, 2026 for gpt-6-luna. Those pages do not say they describe the ChatGPT versions, so the closest answer, as of October 2026, is spring 2026 per OpenAI's API model pages. The cutoff moves with every model. In December 2025, the release notes gave GPT-5.2 a knowledge cutoff of August 2025.

Anything your company changed after the cutoff (a new product, a price, a rebrand) exists for ChatGPT only when it searches. From memory, it answers with the old version. Our guide to fixing stale AI facts about your company covers what to do when that happens.

How does ChatGPT's live web search work?

OpenAI's ChatGPT search help page says ChatGPT "may search the web automatically when your question would benefit from current information." You can also switch search on by hand. Then ChatGPT "typically rewrites your query into one or more targeted queries" and sends them to its search providers. Saved memories can shape that rewrite, and your general location may be shared with the providers; your IP address and account details are not passed on.

Here is how often each engine searched, read and cited in our tracking:

Page counts for ChatGPT cover 14 days (7 to 20 October 2026); for Perplexity and AI Overviews, 7 days (14 to 20 October 2026).

ChatGPT's search rate also moved. It searched in 1,000 of 1,022 answers in the week of 23 September, then 796 of 1,022 in the week of 7 October, a drop from about 98% to about 78%. The drop's start lines up with the dates the Chat tab moved to GPT-6: 7 October for Plus, Pro, Business and Enterprise, 8 October for Free and Go. We can see the timing. We cannot prove the switch caused it. Across the last 14 days, 144 of the 146 questions were searched on some days and answered without a search on others. Search is a per-answer decision, even for the same question, and ChatGPT doesn't give the same answer to everyone either.

You can see the rewriting. In 1,458 of 1,596 searched answers (91.4%), ChatGPT exposed the queries it ran: a median of 2 per answer. Asked for an "AI search marketing agency for SaaS companies," it searched "AI search optimization agency SaaS companies GEO LLM SEO agency." That rewrite uses the market's vocabulary (GEO, LLM SEO), which may not be the words on your site. The mechanics are in our explainer on fan-out queries.

An answer without a search can still carry links. Of 448 unsearched ChatGPT answers, 382 linked to companies while listing no sources. Our data does not show where those links came from.

Does ChatGPT use Bing or Google?

OpenAI does not say. Its help page says ChatGPT search "sometimes partners with other search providers," and the 2024 launch post says it uses "third-party search providers, as well as content provided directly by our partners." The help page lists provider privacy policies: the Microsoft privacy statement and the Shopify privacy policy. It names no search index in the text.

Bing's role is on the record as history. On 23 May 2023, Microsoft announced "Bing is coming to ChatGPT as the default search experience." That was the Browse with Bing era. No current OpenAI page says ChatGPT search runs on Bing's index.

Neither OpenAI nor Microsoft names the index ChatGPT search uses today. The accurate answer for October 2026: OpenAI uses third-party search providers, lists Microsoft's privacy statement among the provider policies, and names no index.

How does ChatGPT choose which sources to cite?

OpenAI's own line is short: "ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information. Placement is not guaranteed." The help page also warns that "Search results and citations can be incomplete, outdated, or incorrect."

What we can measure is the funnel. Across 1,596 searched ChatGPT answers, ChatGPT read 41,508 pages and cited 4,717 of them inline, counting each page once per answer. That is 11.4% of read pages cited. And 555 searched answers (34.8%) cited no page inline at all. Being read is the first filter; being cited is a much narrower second one. Perplexity, over 7 days on the same questions, cited 29.5% of what it read.

For B2B SaaS buying questions, ChatGPT cites company websites. Of those 4,717 cited pages:

• Company and agency sites: 4,299 (91.1%)
• Publishers and editorial sites: 126 (2.7%)
• Other reference sites: 120 (2.5%)
• Unclassified sites: 118 (2.5%)
• Review sites such as G2 and Clutch: 28 (0.6%)
• Other user-generated sites: 20 (0.4%)
• Reddit and forums: 5 (0.1%)
• LinkedIn: 1
• Wikipedia: 0 (retrieved once, never cited)

The other engines lean elsewhere on the same questions. Over 7 days, YouTube made up 7.7% and LinkedIn 6.9% of the pages Google AI Overviews cited, and 39.7% of the overviews that carried sources cited YouTube. For Perplexity, company and agency sites made up 90.9% of cited pages, and Reddit none. Our AI citation benchmark for B2B SaaS breaks source types down across engines, and how to become a trusted LLM source covers why some sites keep getting picked.

Does ChatGPT get its information from Reddit?

Yes, through the open web and through a deal. OpenAI's stated training scope covers public forums, and on 16 May 2024 Reddit announced that "OpenAI will access Reddit's Data API, which provides real-time, structured, and unique content from Reddit."

How much Reddit shows up depends on what you measure:

• Ahrefs Brand Radar (US, all topics, updated 2 September 2026): reddit.com ranks first, "capturing 16.8% of citations," measured as share of top-source mentions.
• Profound, measured on about 730,000 US ChatGPT.com conversations with a citation from October to December 2025: Reddit at 3% of all citations and 13% of conversations with citations, second to Wikipedia.
• Semrush (230,000+ prompts over 13 weeks in 2025): ChatGPT cited Reddit in close to 60% of responses in early August, collapsing to around 10% by mid-September.
• Our B2B SaaS buying questions (7 to 20 October 2026): ChatGPT read a Reddit or forum page 102 times and cited one 5 times, 0.1% of the pages it cited.

Those readings use different units, time windows and question sets, so never average them. Reddit's weight in ChatGPT depends on the question type and swings over time; for vendor-shortlist questions this month, it barely registers.

What ChatGPT does not do

• It does not look everything up live. Without a search, it answers from what the model learned before its cutoff. Search runs when a question needs current information or when you ask for it.
• It does not keep a copy of the web. Training leaves no stored copies to consult.
• It does not treat your memory as a shared source. Memory is per account: "Memory sources are only shown within your account experience." Personal chats may be used to train future models only when "Improve the model for everyone" is on. By default, OpenAI does not train on Business, Enterprise or Edu workspace content.
• It does not intentionally collect paywalled pages when it gathers public web data. Paid content can still reach OpenAI through a licensed deal: Axel Springer's agreement covers content from its brands, "including otherwise paid content."
• It does not show a site in search answers once that site blocks OAI-SearchBot. Per OpenAI, such a site "will not be shown in ChatGPT search answers, though can still appear as navigational links."
• It does not sell answers, per OpenAI. Ads are tested on Free and Go tiers in 2026, and OpenAI says "Ads do not influence the answers ChatGPT gives you." Shopping results, it says, "are not ads, nor influenced by any OpenAI partnerships."
• It is not always right. OpenAI says plainly: "ChatGPT also hallucinates."

How we measured this

Every figure marked as ours comes from LoudFace's own Peec project, read on 21 October 2026. Peec asks each tracked prompt once a day per engine from a US location. The project tracks 150 B2B SaaS agency-buying questions; we dropped the 4 that name LoudFace, leaving 146. A read page is an entry in the answer's sources list; a cited page is one of those entries that the answer links inline. Each page counts once per answer, however many times the answer links it.

The limits: Peec does not show which search provider or index ChatGPT used, which ChatGPT plan or model answered, or whether the session was logged in. Counts of inline-cited pages miss sources shown only in the sources panel. The questions are agency-buying questions, so the source mix describes that question type, not ChatGPT in general; all-topic studies like Ahrefs' and Profound's put Wikipedia and Reddit near the top. Perplexity and AI Overview page data covers 7 days and ChatGPT's 14, so compare per-answer figures, not totals.

What this means for a B2B SaaS brand

For B2B buying questions like the ones we track, ChatGPT will most likely search, read about 22 pages, and cite about 3 of them. Three things follow from that.

Get into the pages it reads. Allow OAI-SearchBot, the crawler OpenAI uses for search inclusion, in robots.txt, and allow requests from OpenAI's published IP ranges. GPTBot is a separate setting for training. OpenAI says "Each setting is independent of the others," so blocking GPTBot does not remove you from search. Then match the words ChatGPT searches with. Its rewritten queries use the market's terms. If ChatGPT searches for "GEO agency" and your site only ever says something else, your page is harder to find.

Then be the page it quotes. ChatGPT cited 11.4% of the pages it read, so most pages on the read list never get the link. Our view at LoudFace is that the page that gets quoted puts a liftable answer at the top: a short answer, a table, a number with its source. A page that never enters the sources cannot be cited at all. Your own pages are half the job. The other half sits off your site: the third-party lists and review sites that name vendors in your category, which ChatGPT reads next to your pages.

Fix what training got wrong. From memory, ChatGPT knows your company as it was before its knowledge cutoff, which OpenAI's API model pages put in spring 2026. Facts you changed since need to be live and crawlable.

For the step-by-step version, read how to get your B2B SaaS recommended in ChatGPT. To see your own position today, LoudFace's free AI visibility audit checks where your brand stands across ChatGPT, Claude, Gemini, and Perplexity versus competitors.