Web Search
Pull current information from the web into chat: Google search, reading individual URLs, source previews, domain filtering via prompt.
The built-in Web Search tool gives the model access to current information from the internet. You enable it with the toggle in the chat input (see Chat overview), and you can use it the same way in Assistants and per-step in Workflows.
What web search looks like from a user perspective
- You enable the web search toggle in the chat (or in an assistant/workflow)
- You ask a question or formulate a research task
- The model formulates its own search queries and searches the web
- Sources are shown with a preview (title, description, favicon)
- The answer is based on the content of the found pages
Two core capabilities
Web Search bundles two closely related functions that the model picks autonomously depending on the task:
Web search
- Method:
builtin-web-search - Inputs:
query,maxResults - Result: hit list with titles, snippets and source URLs
- By default 10 sources per search query, comparable to the first Google results page
Read a URL (website reader)
- Method:
builtin-read-url - Input:
url - Result: cleaned page content in text/markdown form
- Use this when you want to point the model at a specific page: "Read https://example.com/whitepaper and summarize the three key claims."
Things to know before trusting web search
10 sources per search query, not the whole internet. The web search roughly matches the first Google results page. For broader research, chain several targeted searches or build a Workflow with several research steps.
Own queries instead of your wording. The model formulates the search queries itself, based on your question. That can lead to unexpected results when your request is ambiguous. If you want certainty, give concrete keywords in the prompt: "Search for 'GDPR Schrems II implications cloud services 2026'".
Domain filtering (workaround)
There is currently no platform setting that restricts sources to specific domains. If you want to allow only certain sites, instruct the model directly in the prompt:
"Research GDPR compliance, use only heise.de and golem.de as sources."
This works reasonably well in practice but is not a hard security boundary. For truly sensitive research, verify the result manually against the named sources.
Prompting tips for better web search results
| Principle | Description |
|---|---|
| Search-engine perspective | Formulate keywords you would also type into a Google search. Prefer "GDPR data processing agreement obligations 2026" over "Tell me about GDPR". |
| Few examples in the prompt | Few-shot examples can mislead the web search. Better to phrase the task directly. |
| State scope and timeframe | "Q3 2025 quarterly reports from DAX companies" yields better hits than "quarterly reports". |
| Define a fallback | "If no reliable sources are found, say so explicitly and do not speculate." |
| Verify claims | For important statements, add a verification step: "Check the statements above against the found sources and flag anything unsupported." |
Typical use cases
- Quick research on current topics not covered by the model's training data
- Fact-checking against external sources
- Summarizing a single web page (URL as input)
- Source-grounded answers in Assistants for research use cases
- Research steps in Workflows with the per-step web search toggle
Limits
- No platform setting for a domain whitelist (workaround via prompt)
- Paywalled and login-protected pages are usually not accessible
- Very recent events (last few hours) may be missing depending on the search engine's indexing
- JavaScript-heavy pages without server rendering sometimes return only skeleton content