How ChatGPT Picks Its Sources: What Actually Decides Who Gets Cited
ChatGPT chooses its sources through two distinct paths. When a web search fires, the model pulls candidate pages from a search index that blends Bing data with OpenAI's own OAI-SearchBot crawling. It then ranks those pages for relevance and quality and cites whichever ones it actually drew on. If search doesn't trigger, none of that happens, the answer comes straight from training data, and no citations appear.
This guide covers the documented mechanics start to finish: the two answer modes, where candidate pages come from, the crawlers OpenAI runs, the step-by-step pipeline behind source selection, how citations get attached, and why the same question can pull different sources on different runs. Throughout, we draw a clear line. What OpenAI has actually confirmed on one side, what practitioners merely infer on the other.
The two modes: when ChatGPT cites sources and when it doesn't
The single most important thing to know about ChatGPT: it doesn't always search the web. It works in two separate modes, and only one of them actually produces citations.
In training-data mode, the model answers from what it learned during training, a static snapshot with a cut-off date. No live pages are involved, so there's nothing to cite and no way for your website to show up. In search-enabled mode, ChatGPT issues live web queries, pulls in current pages, and attaches inline citations to the sources it used. The user can trigger search by toggling it on, or the model can decide on its own that a question needs fresh or specific information it doesn't already have.
The question of where ChatGPT gets its information really has two honest answers, and which one applies depends on the mode. If you want your site to even be a candidate, you're competing for the search-enabled path. That's the only place where retrieval, ranking and citation actually happen.
| Mode | When it triggers | Where information comes from | Citations shown? | Can your site appear? |
|---|---|---|---|---|
| Training-data answer | Default for general questions the model can answer from memory | The model's training snapshot (static, with a cut-off date) | No | No, no live retrieval occurs |
| Search-enabled answer | When the user enables search, or the model judges a question needs current or specific facts | Live web pages retrieved from search-index data plus OpenAI's own crawling | Yes, inline links to pages used | Yes, if your page is retrievable and selected |
Where the candidate pages come from: index, crawlers and robots.txt
When it runs a search, ChatGPT doesn't crawl the entire web in real time. Instead, it pulls candidate pages from an existing index, publicly, ChatGPT search has drawn on Bing index data alongside OpenAI's own crawling. Google's index isn't part of the mix. OpenAI hasn't laid out the exact blend of third-party index versus first-party crawl data, so any precise split you see elsewhere should be treated as a guess.
OpenAI's own crawling is where most of the confusion starts. The company runs three separate crawlers, each with its own job and its own robots.txt token. Block one, and the other two keep going, site owners who add a single blanket rule often assume they've opted out of everything, and they haven't.
| Crawler | What it does | Affects ChatGPT citations? | robots.txt token to allow/block |
|---|---|---|---|
| OAI-SearchBot | Surfaces and indexes pages for ChatGPT search results | Yes, this is the crawler tied to search visibility and citations | `OAI-SearchBot` |
| GPTBot | Collects publicly available data used to train future models | Indirectly at most, governs training data, not live search | `GPTBot` |
| ChatGPT-User | Fetches a specific page on a user's behalf during a session | Only for that on-demand fetch | `ChatGPT-User` |
The practical takeaway: if you block `GPTBot` to keep your content out of training data but you want to remain citable in ChatGPT search, leave `OAI-SearchBot` allowed. Blocking `GPTBot` alone does not remove you from search citations, because those depend on a different bot and on index data. Conversely, an over-broad `Disallow` that catches every user-agent can quietly remove you from the search-enabled path altogether. Check your robots.txt against the specific tokens above rather than assuming one rule covers all three.
The selection pipeline step by step
Once search kicks off, source selection follows a pattern you can actually observe. Nobody's published the exact internals, but the behavior stays consistent enough that it's fair to describe it as a pipeline:
- Query interpretation and rewriting. ChatGPT takes your prompt and turns it into one or more search queries, often several at once. A single question can split into sub-questions, a pattern people call query fan-out.
- Retrieval. Each query is run against the web index, returning a set of candidate pages.
- Candidate ranking. The candidates get ordered by relevance and apparent quality for the specific question asked. There's no single public "ranking factor" driving this.
- Content extraction. The most promising pages are read, and the relevant passages are pulled out.
- Synthesis. The model composes an answer from the extracted material, blending multiple sources.
- Citation attachment. The pages that actually informed the answer are linked inline.
Site owners should care about query fan-out. One user question splits into several searches behind the scenes, so a page can get pulled in for a sub-question nobody expected it to rank for. That changes what "did I get cited?" even means. It's rarely a clean yes or no; it comes down to which of those fanned-out queries your page actually competed for.
How citations get attached: and how reliable they are
Citations show up as inline links, usually alongside a source list, pointing back to the pages the model actually drew on. In interface terms, "chatgpt citations" are just the retrieved pages ChatGPT decided it used to build its answer, a visible trail leading back to the web.
That trail helps, but it's no guarantee of accuracy. A citation only tells you a page got pulled up and referenced; it doesn't certify that the model summarised it correctly. Synthesis can still misstate things, over-generalise, or blend one source's meaning with another's. If a claim matters (a price, a regulation, a statistic), click through and check it against the cited page instead of trusting the paraphrase. The citation marks where verification starts, not where it ends.
Why citations differ run to run
Ask ChatGPT the same question twice and you might get different sources back. That's expected, not a bug. Generation is non-deterministic, so the model can phrase queries and put together answers differently each run. Retrieval shifts too. Index freshness, re-ranking, small changes in the rewritten queries, any of these can surface a different set of candidates.
Anyone tracking visibility needs to take one thing seriously: don't judge your presence from a single prompt. Track citation share across many runs and phrasings, over time. That way you're reading a pattern, not one roll of the dice. Our guide on how to track AI citations lays out a structured approach to that kind of measurement.
What makes a page selectable: the factors that matter
Here is the honest part. OpenAI publishes documentation for its crawlers, not a ranking-factor list for ChatGPT search. Anyone claiming to know the full algorithm is guessing. What follows is practitioner consensus, the factors that repeatedly appear to help pages get retrieved and cited, clearly separated from the handful of things OpenAI actually documents.
| Factor | Why it plausibly matters | Who confirms it |
|---|---|---|
| Crawl access for OAI-SearchBot | If the search crawler cannot reach the page, it cannot be a candidate at all | OpenAI crawler documentation (confirmed) |
| Front-loaded answers | Extractable, direct answers near the top are easier to lift into a synthesised response | Practitioner consensus (inferred) |
| Clear structure and HTML tables | Headings and native tables make facts machine-readable during extraction | Practitioner consensus (inferred) |
| Freshness | Current pages are favoured for questions where recency matters | Practitioner consensus (inferred) |
| Topical authority and brand mentions | Established, frequently referenced sources tend to be trusted candidates | Practitioner consensus (inferred) |
| Content extractability | Content in plain, accessible HTML (not locked behind scripts or blocked assets) is easier for a bot to read | Practitioner consensus (inferred) |
Only the first row counts as documented fact: reachability by the search crawler. Everything beneath it is evidence-based inference, drawn from how retrieval and extraction actually behave. It's sensible to act on, but it isn't an OpenAI-confirmed formula. Be wary of any pitch promising "guaranteed ChatGPT rankings." Those ranking factors just aren't published. Two supporting building blocks are worth understanding on their own: the llms.txt file and schema markup for AI search.
Understanding ChatGPT versus a standard Google search comes down to how each one picks what to cite. We break that down in how ChatGPT decides which pages to cite versus Google. Perplexity works in much the same way, retrieving sources and citing them, and we've laid out that process too in ChatGPT's equivalent selection process on Perplexity.
What this means for UAE businesses
Buyers in the UAE now reach for ChatGPT before they even open a browser tab, asking it to shortlist vendors, compare options, and explain a category. When ChatGPT answers in search-enabled mode, being retrievable and citable becomes a new visibility layer. It sits alongside classic search rankings rather than replacing them.
At RATH Infotech, we put this into practice on our own site: an llms.txt entry, structured data, HTML-native tables, all built so answer engines can read us cleanly, and all of it verifiable once live. We're upfront about the limits, too. No agency holds client AI-citation metrics worth quoting yet, and no one, including us, can guarantee a citation. What we can do is apply the documented mechanics methodically. We've served Dubai and all of the UAE since 2014, and we treat AI-search readiness as engineering, not promises. See how we apply this for clients on our LLM SEO page.
Frequently asked questions
Does ChatGPT cite its sources?
Only in search-enabled answers. When ChatGPT runs a web search, it shows inline citations linking to the pages it drew on. Answering from training data alone is a different story: there's no live source to cite, so what comes back has no links attached to any web page.
Where does ChatGPT get its information?
Two places, depending on the mode. For general answers, it draws on its training data, a static snapshot with a cut-off date. When search triggers, it pulls from live web pages retrieved through search-index data plus OpenAI's own crawling. Only the second path can include your website.
Does ChatGPT use Google or Bing?
ChatGPT search pulls from Bing's index and OpenAI's own crawler, OAI-SearchBot; it doesn't touch Google's index at all. Nobody's published the exact split between third-party index data and first-party crawling, so any precise breakdown you come across elsewhere should be treated as a guess, not a documented fact.
Does ChatGPT crawl websites like Google does?
Yes, though it comes down to three separate crawlers. OpenAI runs OAI-SearchBot to surface content in search, GPTBot to gather training data, and ChatGPT-User for fetches made on the fly during a session. Each one checks robots.txt on its own. That means you'll need distinct tokens for each bot rather than a single rule covering all three.
If I block GPTBot, will ChatGPT stop citing my site?
Not necessarily. GPTBot handles training-data collection; search citations run on a separate track, tied to OAI-SearchBot and index data. Block one bot and the other keeps working. Want to stay citable in ChatGPT search while opting out of training? Block GPTBot, but leave OAI-SearchBot allowed in your robots.txt.
Why does ChatGPT cite different sources for the same question?
Generation isn't deterministic, and retrieval shifts from run to run, index freshness, re-ranking, and slightly different rewritten queries all move the candidate set around. Don't judge one answer. Track your citation share across many prompts instead. Our guide on tracking AI citations covers how to do that properly.
Can you trust the sources ChatGPT cites?
A citation only shows what got pulled up. It doesn't prove the summary is accurate. The model's synthesis can still misstate a source, or blend two meanings into one. So treat citations as a trail for verification, not proof in themselves: for anything that actually matters, a price, a rule, a figure, click through and check it against the cited page.
How do I get ChatGPT to cite my website?
Short version: be crawlable by OAI-SearchBot, publish clear front-loaded answers, and build genuine topical authority. The full playbook lives in our guide on how to rank in ChatGPT. No one can guarantee citations, including us, the ranking factors are not published, so treat any guarantee as a red flag.
Next steps
Once you've got the mechanism down, the actual work splits three ways: put it into practice, compare it across engines, and measure the results. Start with the source-selection mechanism applied as a playbook in how to rank in ChatGPT to see it in action. Then check ChatGPT's equivalent selection process on Perplexity for the same logic at work. From there, how to track AI citations shows you how to measure your share over time. Want us to handle it instead? Get a free growth audit, and we'll assess your AI-search readiness.