How ChatGPT Chooses Sources: What OpenAI Documents and What It Doesn't

How ChatGPT chooses sources is only partly public: OpenAI documents that ChatGPT rewrites a question into search queries, sends them to partner search providers, reads the results and ranks them using multiple factors meant to find relevant, reliable information. It does not publish those factors, so anything more specific is what website owners observe, not what OpenAI states.

• ChatGPT can turn 1 question into several targeted queries for partner search providers, then run more specific ones after reviewing the first results (OpenAI Help Center on searching the web with ChatGPT).
• ChatGPT search launched on 31 October 2024 on a fine-tuned version of GPT-4o, drawing on third-party search providers and content that partners supply directly (OpenAI's announcement of ChatGPT search).
• OpenAI's developer search tool keeps 2 lists apart: inline citations for the most relevant references, and a sources field with every URL the model consulted (OpenAI's web search tool guide).
• Websites that opt out of OAI-SearchBot are not shown in ChatGPT search answers, and a robots.txt change takes about 24 hours to take effect (OpenAI's crawler documentation).

What does OpenAI say about how ChatGPT picks its sources?

OpenAI describes a search, read and cite process, and says ranking uses multiple factors aimed at relevant, reliable information. It does not list the factors or how much each one counts.

The public record comes from three OpenAI pages, and it is worth reading them side by side. The help page on searching the web says ChatGPT may search automatically when a question would benefit from current information, rewrites the question into one or more targeted queries for its search partners, and may refine those queries after reviewing the first results. The launch announcement adds that ChatGPT search draws on third-party search providers and on content supplied directly by partners, and names 14 news publishers it worked with at launch.

Step	What OpenAI documents	What it leaves out
Deciding to search	ChatGPT searches when a question would benefit from current information, or when the user turns search on	How it judges that a question needs current information
Writing queries	It rewrites the question into targeted queries, adding a general location for local questions and, with memory on, details from saved memories	The exact wording of the queries it sends
Finding pages	Partner search providers and content supplied directly by partners	Which providers are used for which questions
Ranking	Multiple factors intended to find relevant, reliable information; placement is not guaranteed	The factors and their weights
Citing	Responses that use search may include citations, and a Sources panel lists cited sources and other relevant links	Why one page is cited and another is not

The gaps matter. Any claim that ChatGPT prefers a certain word count, schema type or number of headings reports what its author observed, because OpenAI has published nothing that specific. Act on what OpenAI states, and treat the rest as a hypothesis to test on your own questions.

Does ChatGPT cite every page it reads?

No, ChatGPT reads more pages than it cites. OpenAI's own documentation separates the pages consulted from the references shown in the answer.

The clearest statement is in OpenAI's guide to the web search tool it offers developers. It says that, unlike inline citations, which show only the most relevant references, a separate sources field returns the complete list of URLs the model consulted. That guide describes the developer tool rather than the ChatGPT app, so it is not proof of how ChatGPT itself behaves, but it shows how OpenAI's search models are built to separate reading from citing.

The ChatGPT help page points the same way from the user's side. It says responses that use web search may include citations, and that the Sources panel, when available, shows cited sources and other relevant links. A website that appears in the Sources panel was therefore not necessarily used in the reply, and the inline citations beside a sentence are the stronger evidence that your page shaped the answer.

For a business this explains a common puzzle. Your page can be found, opened and read on every relevant question and still never be the one quoted, because another page answered the same point more directly or with firmer evidence. Being read is the entry ticket; being cited depends on what your page gives the model to use once it is open.

Which parts of the choice can your website control?

A website controls whether ChatGPT may read it, whether it matches the queries ChatGPT writes, and how easy its pages are to quote. Ranking itself stays with OpenAI and its partners.

Eligibility is fully in your hands. OpenAI's help page says a website must allow OAI-SearchBot in robots.txt, and its host or CDN must accept traffic from OpenAI's published search bot IP addresses, to be eligible for inclusion. The crawler documentation adds that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links, so a single line in robots.txt or a firewall setting can remove a business from every answer. You can check what ChatGPT's search bot sees with the AI Search Visibility Checker, which reads robots.txt for OAI-SearchBot and asks for your homepage as the bot to catch blocks that robots.txt does not show.

Matching the rewritten queries is partly in your hands. OpenAI's examples show a vague question becoming a precise query with a location or a year added, so pages that use the plain words customers use, including the service, the town and current dates, have more chances to be returned by the partner search. A hypothetical bakery in Leeds that only says "artisan treats for every occasion" gives a query such as "birthday cake order Leeds" little to match, while a page headed "Ordering a birthday cake in Leeds" matches it directly.

Quotability is the part most owners can improve fastest. A passage that names its subject, answers in its first sentence and includes a specific fact, such as your own lead time or service area, can be used as written. OpenAI's help page also tells users to check publication and update dates and to prefer authoritative sources, so a page that shows when it was reviewed and links to the official source behind any rule gives the reader a reason to trust it.

What do business owners observe about the sources ChatGPT cites?

Owners who test their own questions often see directories, comparison lists, review platforms and clear answer pages cited. Treat these patterns as observations, because OpenAI has not confirmed any of them.

Questions that ask for a recommendation, such as "best wedding photographer in Cardiff", tend to be answered from pages that compare several businesses, since a model cannot take one business's word that it is the best. Questions that ask for a fact, such as "how long does a boiler service take", tend to be answered from pages that state the fact plainly in their opening lines. Both patterns fit what OpenAI does say about seeking relevant, reliable information, but neither is a published rule.

Answers also change. The same question asked twice, on different days or by different users, can cite different pages, partly because ChatGPT rewrites queries using location and, where memory is on, saved details about the user. A single test therefore proves little, and the only reliable reading comes from asking a fixed set of questions repeatedly and counting how often each website is cited.

Third-party studies of ChatGPT citations exist, and some are widely shared. They are usually small, use their authors' own question sets, and go out of date quickly as OpenAI changes its models and partners, so none is relied on here. Your own question panel, run monthly on the questions your customers ask, is a better guide for your business than any general study.

Why does ChatGPT cite a competitor instead of your website?

When a competitor is cited and you are not, the cause is usually one of five things: access, findability, a weaker answer, thinner evidence or fewer mentions elsewhere.

Work through them in order, because each one blocks the next. If OAI-SearchBot cannot reach your page, nothing else matters; if it can, but your page never appears in the searches ChatGPT runs, the quality of your answer is irrelevant. Only once both are true does the comparison between your page and the competitor's come down to the content.

What you see	Likely cause	What to check
Your site never appears, even in the Sources panel	OAI-SearchBot blocked by robots.txt, firewall or CDN	Robots.txt rules and how the server answers the bot
Your link appears with only a title	Bot blocked, so ChatGPT shows a navigational link	The OAI-SearchBot group in robots.txt
Your site is listed in Sources but not cited in the text	Another page answered more directly	The first 60 words of both pages
A directory or list is cited for "best" questions	The answer was built from pages comparing businesses	Whether you are listed on the cited pages
Old details about you appear	An outdated page or listing was read	Every listing with your name, address and phone

The comparison of the two pages is where most of the learning happens. Put the competitor's cited page next to yours and read only the first screen of each: whether the answer comes first, whether it names the business and place, whether it gives a figure or a date you can check. The difference is rarely length and usually directness and specificity, which is something you can change within a day.

Can partnerships or ads change which sources ChatGPT chooses?

Content partnerships and ads exist, but neither buys a citation for an ordinary business website. Citations in answers are earned by the page.

OpenAI's launch announcement says ChatGPT search uses content supplied directly by partners, and the named partners at launch were large news publishers. That route is a commercial agreement for publishers whose content is the product, not a programme a local firm can join. OpenAI also does not say how partner content is weighed against other pages, so even for partners it is not a promise of citation.

Ads are a separate product with their own crawler. OpenAI's crawler documentation says OAI-AdsBot checks landing pages submitted as ads in ChatGPT and may use their content to decide when an ad is relevant, and that its data is not used to train foundation models. An ad is a labelled placement, while the help page states that placement in search results is not guaranteed, so there is no published way to pay for a cited source.

The broader question of how to get cited by ChatGPT, from robots.txt to mentions on other websites, rests on the same point. Access, a page that matches the searches customers trigger, and an answer worth quoting are what a business controls. Those three are where the effort belongs, and none of them can be bought.

Questions owners ask about ChatGPT's source choices

Does ChatGPT use Google or Bing to find sources?

OpenAI's help page says ChatGPT sends rewritten queries to partner search providers without naming them on that page. It links Microsoft's privacy statement among those that explain how providers may process queries. Many owners make sure their website is indexed by both Bing and Google, since being findable in ordinary search is a precondition either way.

Does ChatGPT prefer big brands over local businesses?

OpenAI has not said so, and its documentation describes relevance and reliability rather than size. Large brands are named on more websites, which gives them more chances to be read and cited. A local business can compete on local and specific questions, where a page that names the service and the town answers better than a national site.

Do newer pages get cited more often?

OpenAI does not publish a freshness rule for ChatGPT search. Its help page does advise users to check publication and update dates, and ChatGPT searches precisely when a question needs current information. Showing a clear review date on pages whose details change is a sensible step.

Can I see every page ChatGPT read for an answer?

Not fully in the ChatGPT app. The Sources panel shows cited sources and other relevant links when available, which is more than the inline citations but not a complete log. OpenAI's developer search tool does return a full list of consulted URLs, but that is a separate product.

Testing which sources ChatGPT picks for your business

Build a fixed panel of customer questions, run it monthly with search on, and record every inline citation and every website in the Sources panel.

Choose 10 to 20 questions that customers ask before they buy, mixing fact questions ("how long does a kitchen fitting take") with recommendation questions ("best kitchen fitter in Harrogate"). Run each in a new chat with search switched on, and record three things in a spreadsheet: the websites cited inline, the websites listed in Sources, and whether your own website appears in either. Ask from the location your customers are in where you can, because ChatGPT adds a general location to local queries.

After the first month, sort the cited websites by how often they appear. The pages that keep winning tell you what ChatGPT finds useful for your questions: if they are directories, get listed; if they are competitors' service pages, compare their openings with yours; if they are official sources, link to them from your own pages. Repeat the panel and the scan every month, change one page at a time, and judge the results over a quarter.