How ChatGPT search works, step by step
·7 min read
ChatGPT search is the web search built into ChatGPT, and it works as a short pipeline. When a question needs current information, ChatGPT rewrites the prompt into one or more terse search queries, sends them to search providers, reviews the results, may search again with narrower queries, and then writes an answer with clickable citations. A page can be one of those cited sources only if OpenAI's search crawler, OAI-SearchBot, is allowed to reach it.
OpenAI describes these steps in its help article on searching the web with ChatGPT. It does not describe the ranking beyond saying ChatGPT weighs "multiple factors" aimed at relevant, reliable information and that placement is not guaranteed. Treat any list of ChatGPT ranking factors as someone's guess.
How does ChatGPT search work?
ChatGPT search runs in six steps between the prompt and the list of sources.
- ChatGPT decides whether the question needs the web.
- It rewrites the prompt into one or more search queries and may add a year, a place or something it remembers about the user.
- It sends those queries to search providers and gets results back.
- It reviews the results and may send narrower follow-up queries.
- It writes the answer, with citations to the pages it used.
- It shows a Sources list and, for some questions, images, a map or product cards.
Steps 2 and 4 decide what happens to your site. Each query is a separate search, and a page that comes back for none of them cannot be cited, however well it answers the original prompt.
When does ChatGPT use web search?
ChatGPT web search turns on by itself when ChatGPT judges that a question would benefit from current information. Users can also force it by picking Search from the tools menu, or by typing / in the message box and choosing Search. It is available on paid and unpaid plans, and to people who are not signed in. For the user's side of the feature, read can ChatGPT search the internet.
For a site owner, the prompts that matter are the ones that trigger a search. "Best accounting software for a three-person firm in 2026" asks for current information. "What is double-entry bookkeeping" does not, and ChatGPT can answer it from training. An answer written without a search carries no citations, so no page gets a link.
How ChatGPT turns one prompt into several searches
ChatGPT rewrites the prompt into short, targeted queries, then writes new ones after it reviews the first results. Search marketers call this query fan-out, borrowing Google's name for it. OpenAI's own example comes from a biotech researcher asking for the latest on cancer drugs that target CCR8:
First query: CCR8 immunotherapy drug development 2025
Follow-up query: CHS-114 conference 2025The queries read like keyword searches. The question words are gone and the names stay, so your page competes on terse strings like these, not on the conversational prompt.
ChatGPT also adds context the user never typed. The first query gained a year. In OpenAI's location example, a request for good restaurants nearby becomes "top restaurants San Francisco", based on a general location taken from the user's IP address, and saved memories can add preferences such as "vegan".
The follow-up depends on the first results. CHS-114 appears nowhere in the prompt. ChatGPT found the name in the first round of results and searched for it next, so whatever those first pages said shapes the second round.
Google's Search Central page on AI features describes the same mechanism for AI Overviews and AI Mode, which may issue several related searches across subtopics and data sources. The Google AI Mode post covers that side.
Where the ChatGPT search engine gets its results
ChatGPT search draws on third-party search providers, OpenAI's own crawler and content supplied by partners. OpenAI's launch post of October 31, 2024 says search uses third-party search providers and content provided directly by partners, and its crawler documentation covers OAI-SearchBot.
| Source | What it supplies | What you can influence |
|---|---|---|
| Third-party search providers | Results for the rewritten queries | Whether those providers index your pages |
| OAI-SearchBot | OpenAI's crawl of sites for ChatGPT search | Your robots.txt rules and CDN bot settings |
| News and data partners | Weather, stocks, sports, news and maps | Nothing, unless you are a partner |
| Merchants and product data providers | Product details, prices and merchant lists | Your Shopify catalog or a direct product feed |
The help article does not name its search providers, but it sends readers to Microsoft's and Shopify's privacy policies for how providers handle queries. Microsoft runs Bing, so a page Bing has not indexed is probably missing from at least one pool ChatGPT draws on. If you have never opened Bing Webmaster Tools, start there.
OAI-SearchBot is the gate for your own pages. OpenAI says sites that opt out of it will not be shown in ChatGPT search answers, though they can still appear as plain navigational links. To be eligible, allow OAI-SearchBot in robots.txt and make sure your host or CDN accepts requests from the IP addresses OpenAI publishes at https://openai.com/searchbot.json. A robots.txt change takes about 24 hours to reach search. GPTBot is a different crawler that collects training data, and blocking it does not remove you from search.
How ChatGPT picks the pages it cites
OpenAI has not published how ChatGPT chooses which results to read or cite. The documented pipeline only gives the order of the hurdles. A page first has to come back for one of the rewritten queries, which is ordinary search ranking at a provider plus OpenAI's crawl. Then the model reviews the results and decides what to use. OpenAI does not say how much of each page it reads.
Your server logs show part of this. OAI-SearchBot requests are the crawl that makes pages eligible. ChatGPT-User requests are pages fetched because someone's question in ChatGPT called for them, which is as close as you get to watching ChatGPT read a page. OpenAI says ChatGPT-User plays no part in deciding what can appear in search. The AI bot log analyzer sorts a pasted access log by crawler, lists the pages each one read and flags visits that use a crawler's name from an IP address its operator does not publish.
How sources and citations appear in ChatGPT search
Answers that used web search can include inline citations, and a Sources button lists the cited pages along with other relevant links. Selecting a citation opens the page, and on desktop web, pointing at one shows a preview. Images link to their source, and on iOS and Android some answers include a map.
OpenAI warns users that results and citations can be incomplete, outdated or wrong. If ChatGPT cites your page for a price you changed last month, the reader who clicks finds the new number and has to guess which one is right.
Shopping and local results in ChatGPT search
Shopping is the one place OpenAI lists what ChatGPT weighs. For questions with shopping intent, ChatGPT shows product cards with images, details and links to sellers, and its shopping help article says those results are chosen independently and are not ads. The inputs it names are structured metadata such as price and description from first-party and third-party providers, the answer the model wrote before it looked at new search results, and OpenAI's safety and product policies.
When someone opens a product, ChatGPT may list merchants ranked on availability, price, quality and whether the merchant is the maker or main seller. Shopify stores are already connected through Shopify Catalog, and other merchants can apply to send OpenAI a product feed. For a store, the product data you send shapes what the card says, so a wrong price in the feed becomes a wrong price in ChatGPT.
Local answers use an approximate location from the IP address, or the device's location if the user shares it, and restaurant answers can show reservation times from OpenTable, Resy or Yelp.
What query fan-out means for your pages
Every query ChatGPT writes is a separate search, and your page has to come back for each one on its own. A prompt like "which CRM should a five-person design agency use" could fan out into searches like these, which are an illustration rather than a logged ChatGPT run:
best CRM for small design agencies 2026
HubSpot vs Pipedrive pricing small team
CRM with client portal for agenciesA page built around "best CRM for agencies" has a chance at the first search and little at the other two. A page with a section comparing the two tools' prices and another on client portals has three chances to be read. Ranking for what a person types into Google tells you little about the strings ChatGPT types.
Coming back in the results is half the job. Getting quoted once ChatGPT reads the page depends on crawler access, an answer present in the raw HTML and sentences specific enough to lift. How to get cited by ChatGPT covers each one, including the robots.txt group for OAI-SearchBot.
Check which ChatGPT searches your page answers
The Fan-out Coverage Checker takes a query and a URL and shows which of the searches behind that query your page answers. It asks ChatGPT the query with web search on and records the searches ChatGPT ran, then adds the related, comparative, recent and implicit searches a model predicts. A model then judges each search against your page's sections by meaning, not exact words, and marks it answered, partial or missing. The coverage score counts answered searches plus half the partial ones.
The report also says whether ChatGPT's answer cited your site and, if not, which sites it cited instead. The checker reads the HTML your server sends, so text that appears only after JavaScript runs is not checked. ChatGPT's answers change between runs, so treat one run as a sample. A run costs 95 credits.