Where ChatGPT actually gets its answers, and what it means for your business
Andrei Nekrasov
TL;DR
- Most ChatGPT answers never touch the live web at all; they come straight from what the model memorized during training, so fixing your website does nothing for the questions that never leave that memory.
- Nobody outside OpenAI, Anthropic, and Google knows what is actually in a model's training data today, which makes most of the "AI is trained on X%" numbers in marketing content unverifiable.
- No single list of top domains covers what AI actually cites. What a model finds depends on the question, so a general ranking of Reddit, Wikipedia, and LinkedIn says little about your own case.
- A page's position in ChatGPT's own search results predicts a citation far better than the page's domain authority does, which means a small site with no backlinks can still get cited regularly.
- The set of sources behind an AI answer can look completely different a day later, so a single check tells you almost nothing about what AI actually cites.
How often ChatGPT searches the web for an answer depends on who measures it: one tracker found 14%, another found 46%. Neither is wrong; they are measuring a system that nobody, including the company that built it, has fully documented in public.
I went through what is actually known about how ChatGPT builds an answer, in research published this year. A citation, the one part you can actually see, turns out to be the smallest part of the story. Most of what shapes an answer never produces a citation at all. Part of it comes from a paid deal nobody discloses. The rest can look nothing like it did a week before.
None of this behavior has settled into anything resembling an algorithm yet. That is bad news if you wanted a checklist. It is good news if you are small and willing to move faster than whoever is currently ahead of you.
Section 1
An AI answer is built from up to 4 different sources
Ahrefs found that a ChatGPT answer can draw on 4 different sources:
- Training data, always in play, since it is the model's base knowledge.
- Live web search, switched on only for some questions.
- Licensed publisher content, from paid deals OpenAI has with more than 20 publishers, with nothing in the citation marking a source as licensed.
- The user's own context, files or memory someone has actually added to the conversation.
A citation is the only one of these 4 you can actually see. The other 3 leave no trace at all. In most answers, they do more of the work than the citation does.
So a list of domains an AI engine cited is never the whole picture. Judging your "AI visibility" by citations alone misses most of what actually shaped the answer.
Section 2
Most ChatGPT answers never touch the live web; they come straight from what the model memorized during training
Not every question even reaches the live web. Most stay inside the model's memory instead.
Ahrefs tracked 66.6 million free ChatGPT responses over August and September 2026 and found that live web search triggered in 14% of them, leaving 86% built only from training memory. A separate tracker that counts only the explicit search button, not any internal search call, puts usage at 35% in February 2026, down from 46% in late 2024. otterly.ai, which tried to estimate the same thing without access to OpenAI's own numbers, lands on a 20% to 35% range and says plainly that it is a guess, not a measurement.
Share of ChatGPT answers involving a live web search, by measurement
Sources: Ahrefs, otterly.ai
These numbers disagree because they measure different things: one counts any internal search call, the other counts only a visible button. They agree on one point: a live web search is the exception, not the rule.
In practice, rewriting a page for AI often changes nothing, not because the rewrite was bad, but because most questions in your category never reach the part of ChatGPT that searches the web at all.
Section 3
Nobody outside OpenAI, Anthropic, and Google knows what's actually in the training data today; the only public breakdown is from 2020
Training data is the one source every answer can fall back on. For most questions, it is the only source in play. The last time anyone outside OpenAI even got a description of what is actually in it was for GPT-3, in 2020.
Filtered Common Crawl made up 60% of it, outbound Reddit links made up 22%, 2 book collections together made up 16%, and Wikipedia made up 3%. That is a 2020 model, not the ChatGPT answering questions today. OpenAI, Anthropic, and Google have not published an equivalent figure for any model since.
So you cannot check any specific claim about what trains today's ChatGPT, no matter how precise it sounds. A line like "Wikipedia is 22% of ChatGPT's training data" even gets its own numbers wrong: the 2020 breakdown puts Wikipedia at 3%, and 22% belongs to the Reddit-linked dataset instead. Treat training-data percentages in marketing content as guesses, not facts.
Section 4
There is no universal list of top domains that AI cites
What a model finds depends on what you actually ask it.
Researchers at the University of Toronto compared Google's search results against 4 AI engines on the same questions. On everyday product questions, Google's sources split roughly evenly between independent media, social platforms and forums, and brand-owned pages: 41%, 34%, and 26%. The 4 AI engines barely touched social platforms at all.
| Engine | Independent media | Social and forums | Brand sites |
|---|---|---|---|
| 41% | 34% | 26% | |
| Claude 4.5 | 65% | 1% | 34% |
| GPT-4o | 57% | 8% | 35% |
| Perplexity Sonar Pro | 50% | 39% | 11% |
| Gemini 2.5 Flash | 46% | 46% | 8% |
A general-interest ranking of AI's favorite domains, the kind that puts Reddit, Wikipedia, and LinkedIn on top, mixes every topic together into one ranking. A dental practice sees the opposite pattern: ChatGPT cites dental practice websites for 54% of the sources in answers about dentists, and national directories for only 6%.
A general top-domains list tells you what gets cited on average, across every subject at once. It does not tell you what gets cited for your specific type of question.
Section 5
A page's position in ChatGPT's own search results predicts citation better than its domain authority
Before ChatGPT answers a question, it runs its own search and gets back a ranked list of pages, much like a results page nobody sees. A page's position on that hidden list turns out to matter more than almost anything else.
Kevin Indig and AirOps tracked 353,799 pages through ChatGPT's own retrieval step and matched each one to its citation rate. ChatGPT cited the page in position 1 of that hidden list 58% of the time; it cited the page in position 10 just 14% of the time. Pages ChatGPT cited on every run did not have more backlinks or a stronger domain than pages it never cited at all.
Citation rate inside ChatGPT's own search, by rank
Source: Kevin Indig and AirOps
A small site with no backlink history can still get cited often, as long as its content actually answers the question well. Domain authority has little to do with it.
Section 6
Getting a reliable picture of what AI cites takes dozens of repeated checks
Run the same question through the same engine a day apart, and the sources behind the answer barely resemble each other.
A June 2026 study tracking 530,875 citations across 4 engines found that 69% of the sources behind a typical answer change from one day to the next. Gemini rebuilds 88% of its sources daily, ChatGPT 79%, Google AI Mode 76%, and Perplexity 44%. Almost none of it survives a full week: the share of sources still present after 7 days ranges from 11% on Perplexity down to well under 1% on Gemini.
Share of sources that changed or stayed, by AI engine and by time span
Changed day to day
Perplexity
Google AI Mode
ChatGPT
Gemini
Survived all 7 days
Perplexity
Google AI Mode
ChatGPT
Gemini
Source: GetMentions.ai
Reddit gives a live example of how quickly this can change. In the middle of August 2026, Reddit blocked every crawler from its entire domain, Google included, through a blanket rule in its robots.txt file. In the same week, Reddit's share of citations on ChatGPT fell 86%, while its share on Perplexity rose 86%, because Perplexity reaches Reddit's content through a separate paid agreement that the crawler block does not touch.
Share of ChatGPT answers with zero sources, by day
Source: higoodie.com
A screenshot of what ChatGPT cited today says almost nothing about what it will cite tomorrow for the same question. If you are judging AI visibility from one check, you might be judging noise.
Section 7
Pages from publishers with a paid content deal get cited nearly 50% more on ChatGPT
Some of what looks like earned visibility is paid for, while nothing in the citation says so.
otterly.ai tracked 129.3 million citations across 7 AI platforms in June 2026 and matched them against 91 confirmed licensing deals covering 314 publisher domains, mostly major news and media outlets. Pages from publishers with an OpenAI deal collected 10.2 citations each on average; pages from publishers without one collected 6.9, a 48% gap on ChatGPT specifically.
Citations per page on ChatGPT, by whether the publisher has a licensing deal with OpenAI
Source: otterly.ai
So part of what looks like earned visibility on ChatGPT is really a paid arrangement with OpenAI, and nothing in the citation marks the difference. If a major media outlet keeps out-citing you on your topic, a paid deal is often the real reason behind it.
Section 8
Up to 18% of the links AI cites as sources do not even open
A citation next to a claim shows that a link exists. It does not show that the link supports the claim.
Researchers at the University of Pennsylvania checked citation links from 10 AI models and research agents against 2 large benchmarks. The weaker models invented 3% to 13% of their cited URLs outright and simply failed to resolve 5% to 18% more. The worst performer, Gemini 2.5 Pro running in its deep-research mode, invented 13% of its links and left 18% broken.
| Model or agent | Non-resolving | Hallucinated |
|---|---|---|
| Claude 3.5 Sonnet, search | 8% | 3% |
| Gemini 2.5 Flash, search | 5% | 5% |
| Gemini 2.5 Pro, search | 6% | 5% |
| GPT-4.1 | 5% | 5% |
| Claude 3.7 Sonnet, search | 9% | 3% |
| GPT-4.1 mini | 7% | 7% |
| OpenAI deep research | 10% | 4% |
| GPT-4o mini, search preview | 9% | 9% |
| GPT-4o, search preview | 9% | 9% |
| Gemini 2.5 Pro, deep research | 18% | 13% |
Source: Rao, Wong, and Callison-Burch, University of Pennsylvania
If an AI answer cites your page next to something inaccurate about your business, the inaccuracy may not come from your page at all. Check whether the link actually opens and says what the AI claims it says. A meaningful share of these links do neither.
Closing thought
One thing explains almost everything above. ChatGPT does not look an answer up in a stable index the way Google ranks a page once and reuses that ranking for months. It rebuilds each answer from scratch, for that one question, every single time someone asks.
That habit changes what matters for your page. A page with quality, relevant answers now matters more to AI than domain authority or years of backlinks.
If you make relevant, useful content and update it in a timely way, chances are, AI will like you.