Gemini web search

Gemini Web Search & Fetch

On-demand web research from the terminal using the Gemini CLI. The CLI exposes two built-in web tools that the agent invokes automatically from a natural-language prompt:

• google_web_search — retrieves up-to-date information via Google Search and returns a grounded summary with source citations.
• web_fetch — retrieves and processes content from specific URLs you provide in the prompt.

When to Use

• You need information newer than the model's training cutoff (latest releases, recent news).
• You're researching the current version of a library, API, framework, or tool.
• You're hunting for solutions to a recent bug, error message, or security vulnerability.
• You found a specific page (docs, blog post, issue) and want it read, summarized, compared, or applied to your code.

Prerequisites

• Gemini CLI installed and authenticated. Verify with:

  gemini --version
  

• An active internet connection.

Capabilities

Tool	Argument	Returns	Best for
google_web_search	query (string)	Synthesized summary grounded in search results, with source URIs and titles	Open-ended research, "what's the latest…", finding relevant pages
web_fetch	prompt containing up to 20 http(s) URLs + instructions	Synthesized response from the page content, with source attribution	Reading/summarizing a known page, comparing multiple sources, extracting details

Procedure

Run a web search

Pass a natural-language prompt; the CLI decides to call google_web_search when current information is needed.

gemini -p "Search for the latest stable release of Bun and summarize the key changes."

gemini -p "Find recent solutions for the Next.js 'hydration mismatch' error and list the top fixes."

The response is "grounded": it includes citations (source URIs and titles) so you can verify the facts rather than trusting an unsourced answer.

Fetch and process specific URLs

Include one or more URLs (up to 20) in the prompt along with what you want done. The CLI calls web_fetch.

gemini -p "Read https://example.com/blog/fixing-memory-leaks and explain how to apply it to a Node.js service."

gemini -p "Compare the pagination patterns described at https://api.example.com/v1/docs and https://api.example.com/v2/docs."

web_fetch triggers a confirmation dialog showing the URLs before retrieval. In Plan Mode it always requires explicit confirmation due to the security implications of accessing external addresses.

Combine search → fetch → apply

A common workflow:

1. Search for an approach: gemini -p "How do I implement auth with Supabase? Search for current guidance."
2. Fetch the authoritative source: gemini -p "Read https://supabase.com/docs/guides/auth and extract the setup steps."
3. Apply it to your project in the same interactive session: "Now create an auth.ts file using that pattern."

Notes & Limits

• google_web_search returns a summary of results, not raw pages — use web_fetch when you need the full content of a specific URL.
• web_fetch accepts a maximum of 20 valid http:// / https:// URLs per prompt.
• web_fetch uses the Gemini API's URL context for retrieval and falls back to fetching raw content directly from your machine if API access fails.
• Always review the cited sources for accuracy before acting on results, especially for security-sensitive or fast-changing topics.