Gemini Often Won't Search the Web — and Won't Tell You It Didn't

Gemini, Google's AI-powered assistant, often fails to search the web when it should, and does not warn the user that it did not search. This issue is apparent when Gemini is asked to provide information that requires current data, such as the weather in a specific location. In such cases, Gemini may confidently provide an answer based on its training data, which may be outdated, without ever searching the web for a fresh update.

There is no indication within the response that a search was not conducted. The reason behind Gemini's selective use of the web search tool is a routing mechanism employed by Google. This mechanism determines, on a case-by-case basis, whether the model should be granted access to its web-search tool. In many instances, Gemini is denied the opportunity to search or is explicitly instructed not to use the tool, even when it is needed.

This routing decision is made before the model begins generating its response, leading to the assistant providing stale answers without any warning. The issue has been documented in Google's own help community, as well as in a leaked Gemini system prompt, and can be consistently reproduced by users. A leaked system prompt explicitly instructs Gemini not to issue search queries for certain prompts, further supporting the theory that the assistant selectively uses its web search tool.

The behavior exhibited by Gemini has been observed across multiple languages, indicating that it is not an isolated issue but rather a deliberate design choice made by Google. This selective use of the web search tool raises concerns about the reliability of Gemini as a trusted source of up-to-date information. While Google argues that the search classifier is an optimization that helps spend resources more effectively, the lack of transparency and user control over this process leaves users questioning the accuracy and timeliness of Gemini's responses.