Recorded prompt
“How does ChatGPT Search choose the sources it cites?”
- Product
- ChatGPT
- Surface
- ChatGPT web with web search
- Model
- ChatGPT automatic model (model version not exposed)
- Locale
- en-US (English answer; zh-CN interface)
- Observed answer
- Open session ↗
- Observed
- 15 November 2026, 21:31 CST
- Fresh conversation
- Yes
What the answer said
ChatGPT Search selects sources based on relevance, reliability and usefulness to the question. It doesn't simply copy the first few results from a traditional search engine. OpenAI confirms that its search system uses multiple factors to identify relevant and reliable sources, although it hasn't publicly disclosed the exact ranking formula or the weight given to each factor. OpenAI Help Center +1 Here's how the process works. 1. Understanding the question ChatGPT first interprets what you're asking and identifies the information needed. It may rewrite your question into several more targeted search queries. For example, if you ask: "What are the benefits of solar energy in the UK?" It might search for: Solar energy benefits. UK solar energy statistics. Government research on renewable energy. Recent studies on solar energy costs. OpenAI Help Center 2. Finding relevant sources ChatGPT can use search providers and information supplied directly by publishing partners to find relevant webpages. It looks for material that helps answer the specific question. Potentially useful sources include: Government websites. Academic research and scientific publications. Established news organisations. Specialist industry publications. Relevant company websites and technical documentation. No particular category is automatically guaranteed to appear. 3. Evaluating the information Search results are ranked using multiple factors intended to identify relevant and reliable information. Some important practical considerations when assessing sources are: Factor What it means Relevance Does the page directly answer the question? Reliability Is the information supported by credible evidence? Freshness Is the information current enough? Specificity Does the source provide useful detail? Accessibility Can the search system retrieve the content? OpenAI hasn't confirmed that these are individual ranking signals with specific weights. They are useful ways to understand what makes a source valuable, rather than a published ranking formula. 4. Attaching citations Once ChatGPT has gathered information, it generates an answer and attaches citations to relevant sources. Ideally: Each citation supports the statement beside it. Different sources support different parts of the answer. Readers can open the original source and verify the information. A citation doesn't necessarily mean that OpenAI endorses the publisher or that every claim on that website is correct. 5. Can websites influence whether ChatGPT cites them? Yes, to some extent. OpenAI provides a crawler called OAI-SearchBot, which helps websites become discoverable in ChatGPT Search. Website owners can: Allow OAI-SearchBot to access their content. Publish clear, useful and relevant information. Keep time-sensitive information updated. Make important information accessible to search systems. However, allowing the crawler doesn't guarantee that a website will be cited. OpenAI Help Center What is the most important thing to understand? ChatGPT Search is not a simple popularity contest. A website isn't guaranteed to be cited because it: Has the most visitors. Has the highest Google ranking. Belongs to the biggest company. Publishes the most content. What matters is whether the retrieved information is useful for answering the particular question. One important limitation: OpenAI acknowledges that search results and citations can sometimes be incomplete, outdated or incorrect. It's worth opening the original sources when accuracy matters. OpenAI Help Center +1 Official reference: OpenAI – Searching the web with ChatGPT.
1 recorded sources
What this observation cannot show
Authenticated free-account ChatGPT browser run in a fresh conversation. The product did not expose its complete retrieval configuration. The source panel reported 1 sources and 1 unique source records were captured. Results can vary by session, locale, account state and date.