Researchers analyzed the efficiency of searching the web with popular LLM

A group of scientists published a study dedicated to analyzing how modern conversational AI agents use web search to respond to user requests. The study compares the strategies of four platforms: ChatGPT, Claude, Grok, and DeepSeek, using both real user data and controlled experiments via API.

The results showed that the frequency of accessing search engines varies greatly between models and does not always correlate with the quality of the final answer. Researchers found that agents apply different strategies for forming search queries, and the built-in search engines of each platform have a pronounced preference for certain domains when providing results.

Particular attention in the study is paid to the problem of data attribution. Despite the fact that answers in most cases are based on found information, some statements of the models are based on data that were not explicitly cited in the final text. This creates risks for the reliability and transparency of system operation when providing factual information.

The authors emphasize that current mechanisms for integrating search into LLM remain insufficiently studied from the point of view of the agent's life cycle. The obtained data indicate the need to revise the architecture of search tools to minimize hallucinations and increase the accuracy of source references.

The work is published under the Creative Commons Attribution 4.0 license. The conclusions of the study can be used by developers to optimize the decision-making algorithms of agents when searching for relevant information on the internet.