Anthropic Enhances Its Claude Chatbot With Web Search Functionality

Claude, the AI-driven chatbot from Anthropic, can now browse the web — a feature that had previously evaded it.

Anthropic announced on its blog that web search is now available in preview for paid Claude users in the U.S., with support for free users and other countries coming soon. From the Claude web app, users can enable web search in their profile settings, and Claude will automatically conduct searches across sites to inform specific responses.

Currently, web search is only functional with the latest Anthropic model driving Claude, specifically Claude 3.7 Sonnet, according to Anthropic.

“When Claude includes web-based information in its responses, it offers direct citations for easy source verification,” the company stated in its blog entry. “Claude processes and presents relevant sources in a conversational format, eliminating the need for you to find search results on your own. With this improvement, Claude’s broad knowledge base is supplemented with real-time insights, offering answers grounded in more up-to-date information.

In my brief testing of the feature, web search didn’t consistently trigger for current events-related questions. But when it did, Claude indeed delivered an answer with inline citations.

With its web-searching capability, Claude reaches feature equivalence with many competing AI-driven chatbots, such as ChatGPT from OpenAI, Gemini from Google, and Le Chat from Mistral. Previously, Anthropic argued against it by stating that Claude was “designed to be self-contained.” It is certain that competitive pressure played a role in the change of direction.

Of course, there is a risk that Claude hallucinates or misquotes web sources. This is a problem for other chatbots. A recent study by the Tow Center for Digital Journalism found that well-known chatbots, such as ChatGPT and Gemini, give wrong answers to over 60% of inquiries. Another report by The Guardian discovered that ChatGPT Search, which is centered around search functionality, can be tricked into producing entirely erroneous summaries.