When ChatGPT searches the web, it turns your question into one or more search queries. It then pulls candidate pages from search results and licensed partners, reads them, and cites the few that answer each part of the question most clearly and credibly. So how ChatGPT chooses its sources comes down to two filters. First, your page has to be reachable and retrieved. Second, it has to be the clearest, freshest and most trustworthy answer among the pages that were retrieved.
The final ranking happens on OpenAI's servers, and nobody outside can see all of it. Still, OpenAI's crawler documentation and a few network studies show enough of the process to act on. This guide covers where answers come from and how a page goes from fetched to cited. It also gives you a grid to diagnose why you are not cited and a checklist you can run on your own site this week.
Where a ChatGPT answer comes from: 4 origins
Not every answer comes from the live web. Each one draws on up to four sources of information, and only one of them gives you a clickable citation.
| Origin | What it is | Can a site owner influence it? |
|---|---|---|
| Training data | What the model learned before release. OpenAI's GPTBot collects content for training (OpenAI). | Slowly and indirectly. No link to you in the answer. |
| Live web search | Pages retrieved at the moment of the question. OAI-SearchBot is the crawler behind ChatGPT search (OpenAI). | Yes. This is the main lever. |
| Licensed content | Publisher content OpenAI has access to. A network analysis found these sources tagged "labrador", including Reuters, WSJ and Wikipedia (Search Engine Journal). | Only by being mentioned in those publications. |
| User context | Files, earlier messages and instructions the user gave. | No. |
Live search is where a business site can win a citation, so the rest of this guide focuses on it. If your goal is to show up in AI answers week after week, our page on getting your pages into ChatGPT and AI answers explains how a steady publishing rhythm feeds that process.
When ChatGPT searches the web, and when it doesn't
ChatGPT does not search for every question. In one test, 3 out of 10 news questions triggered no web search at all (Suganthan). The model answered from what it already knew, so no site got a citation.
In practice, search is more likely when the answer depends on something current or specific: prices, rules that change, comparisons, recent releases, or a named tool or place. Stable definitions are more often answered from memory.
What this means for you: aim your pages at questions where a good answer needs fresh, specific facts. "How much does a bookkeeping service cost per month in 2026" gives ChatGPT a reason to search. "What is bookkeeping" mostly does not.
How one question becomes many searches (fan-out)
When ChatGPT does search, it rarely runs your question as typed. The same network study found that a complex question in thinking mode can fan out into 15 to 40 sub-queries, while the fast model sends just one reworded query (Search Engine Journal).
Take an illustrative example. The owner of a small accounting firm asks: "What's the best way for a 5-person accounting firm to get clients from Google?" Behind the scenes, the sub-queries could look like this:
- seo for accounting firms
- google business profile for accountants
- how much does seo cost for a small business
- blog topics for an accounting firm
- how long does seo take to work
Each sub-query returns its own results. A page that answers the main question and its natural sub-questions, each under a clear heading, has more chances to be retrieved for at least one of them.
Fetched, cited and mentioned are not the same
These three words describe three different outcomes:
- Fetched: ChatGPT downloaded and read your page while building the answer.
- Cited: your page appears as a source link the user can click.
- Mentioned: your brand is named in the answer text, with or without a link.
The gap between fetched and cited can be huge. In the network analysis, Reddit was fetched 278 times but cited 11 times, and YouTube was fetched 201 times with zero citations (Search Engine Journal). The sample was about 1,240 sources, so read it as a direction, not a rule.
The lesson still holds. Getting read is only step one. Your page also needs a passage that ChatGPT can lift and attribute with confidence.
How ChatGPT picks which websites to cite
Once pages are retrieved, the model has to pick a handful. Zapier looked at how ChatGPT explains its own choices. The criteria it lists are precise keyword matches, search intent, recency, credibility, trustworthiness and a variety of perspectives. Zapier also reports that the AI leans mostly on roughly the top 20 search results (Zapier). That comes from the model's own account, so treat it as a guide rather than proof.
Here is what each criterion looks like on a real page:
- Relevance to the exact sub-query. The heading and first sentences match the question in plain words.
- Freshness. A visible update date, current figures, and no outdated advice.
- Credibility. A named author or business, real contact details, and an About page.
- Clarity. A short answer that makes sense on its own, without the paragraphs around it.
- Agreement with other sources. Claims backed by linked sources. A figure that no other page confirms is easier to skip.
- Variety. ChatGPT often mixes source types: a forum thread, a publisher and a brand site. Your site is competing for one seat at the table, not all of them.
The 3 OpenAI crawlers and your robots.txt
OpenAI runs three bots, and each one follows its own rules (OpenAI):
| Bot | Job | What blocking it does |
|---|---|---|
| OAI-SearchBot | Finds pages for ChatGPT search | Your site is not shown in ChatGPT search answers (navigation links may still appear) |
| GPTBot | Collects content for model training | Your content is not used for training. It does not remove you from search. |
| ChatGPT-User | Visits a page when a user asks ChatGPT to | It does not crawl on its own, and robots.txt rules may not apply |
A robots.txt that allows search but blocks training looks like this:
User-agent: OAI-SearchBot
Allow: /
User-agent: GPTBot
Disallow: /
OpenAI says changes take about 24 hours to reach its systems. Not sure what your current file allows? The free SEO checker tests whether OAI-SearchBot, Googlebot, Bingbot and PerplexityBot can reach a page, and whether it can be indexed. For the full setup, see how to rank in ChatGPT.
Why ChatGPT cites some sites and not others: a diagnostic grid
Use this grid after you have tested your own questions in ChatGPT (step 7 of the checklist below).
| What you see | Possible cause | What to do |
|---|---|---|
| Your site never appears, on any question | OAI-SearchBot blocked, page not indexed, or text only loaded by JavaScript | Check robots.txt, your sitemap and indexing first |
| Your brand is named but competitors get the link | Your page has no short, quotable answer to that exact question | Add a direct two-sentence answer under a matching heading |
| Directories or review sites are cited about your own business | ChatGPT may favor independent sources next to brand sites | Keep your listings on trusted directories complete and accurate |
| Cited last month, gone this month | Citations move a lot between runs | Update the page and keep publishing on the topic |
| The answer has no citations at all | No web search was triggered | Target questions that need current or specific facts |
| An old page of yours is cited instead of the newer one | The old page may have more links and a longer history | Update the old page, or link it clearly to the new one |
A 7-step checklist to give your site a fair chance
- Allow OAI-SearchBot in robots.txt. Decide on GPTBot separately, based on whether you accept training use.
- Get indexed. ChatGPT pulls from web search results, so a page no search engine has indexed is less likely to be retrieved. Submit your sitemap and fix pages that are not indexed. If some are stuck, see why blog posts are not indexed by Google.
- One page per real question, with the sub-questions from the fan-out under their own H2 and H3 headings. To find those sub-questions, type your main topic into the free keyword generator and turn the closest real searches into headings.
- Open each section with a standalone answer of one or two sentences, then add the detail.
- Source your facts with links and keep a visible update date.
- Earn mentions where ChatGPT reads: industry directories, local or trade press, and honest answers on forums where you have real expertise. No spam. It gets removed, and it hurts trust.
- Test it yourself. Write down 10 questions your customers actually ask. Run them in ChatGPT with search on, once a month, and record for each one whether you were cited, mentioned or absent. Then track ChatGPT referrals in GA4 to see if citations bring visits.
The method works whatever kind of site you run. A service business should test local, price and "who should I hire" questions. A SaaS company should test comparison and how-to questions about its category. An agency can run the same 10-question test for each client and report it monthly.
What nobody can tell you yet
Be careful with anyone who claims to know the exact formula.
- The ranking is opaque. It happens on OpenAI's servers, and studies only see what is fetched and shown.
- Answers change. They vary between runs, between users and after model updates.
- Samples are small. The best network study used one Pro account and roughly 1,240 sources, mostly SaaS and tech questions (Search Engine Journal).
- No one can guarantee a citation. What you can control is being reachable, indexed, clear and sourced on the questions your customers ask.
How AI SEO Crew helps your pages get picked
Everything above leads to the same work. You publish pages that each answer a real question your customers ask. You back them with sources. You get them indexed. Then you keep doing it, month after month. Doing it once barely moves anything. A steady flow of pages on real questions gives ChatGPT more chances to retrieve one of yours when it searches.
That is the work the crew takes off your hands:
- Scout reads your site and finds the real Google searches around your business, the questions a fan-out search is built from.
- Strategist picks the searches that bring customers, not just traffic.
- Writer researches each topic on the web, then writes the article with its sources.
- Editor checks every fact and link and adds internal links, so no claim goes live unsupported.
- Publisher posts the article to your site, sends your sitemap to Google and pings IndexNow, so new pages can be found and indexed quickly.
Every article is written to be quoted by ChatGPT, Perplexity and Google's AI Overviews. The LLM SEO page shows how the crew handles it, and our AI SEO guide covers the full strategy. No tool can guarantee a citation, and we can't either. What the crew gives you is more pages that are reachable, indexed, clear and sourced: up to 50 articles a month for $49.
Start free and see your SEO plan in a few minutes, no card needed.
Frequently asked questions
Does ChatGPT use Google or Bing to find sources?
OpenAI does not publish the full list of search providers it uses. A network analysis published by Search Engine Journal saw sources tagged as open web results, licensed publisher content and third-party scraping services. The safe move is to be well indexed in the major search engines and to allow OAI-SearchBot.
Do I need to allow OAI-SearchBot in robots.txt?
Yes, if you want to appear in ChatGPT search answers. OpenAI states in its crawler documentation that sites that opt out of OAI-SearchBot are not shown in those answers. Blocking GPTBot is a separate choice that only affects training.
Why does ChatGPT sometimes answer without citing anything?
Because it did not search the web for that question and answered from its training data. In one published test, this happened even on some news questions. Questions that need current or specific facts are more likely to trigger a search and citations.
Does Reddit get cited more than brand sites?
Reddit gets fetched often, but in one study published by Search Engine Journal it was cited far less than it was read: 278 fetches for 11 citations. A clear, sourced page on your own site can still win the citation for a specific question. Being mentioned in community threads helps your brand appear in the answer text.