Google Launches Deep Research Tool for Gemini Advanced

Google has launched Deep Research, a new feature for its Gemini Advanced subscription that allows the AI to autonomously browse the web and compile reports on complex topics [1]. The move signals a shift toward "agentic" AI, where models perform multi-step tasks on behalf of users rather than simply answering single-turn prompts [1].

Automating the research process

Deep Research is designed to replace the manual process of opening multiple browser tabs to hunt for information [1]. When a user submits a query, the tool generates a research plan, executes web searches, and compiles the findings into a report that includes clickable citations [1]. The output can be exported directly to Google Docs, providing a workflow integration for users conducting industry analysis or academic research [1].

The tool operates within the existing Google Search architecture, meaning it adheres to established safety policies regarding restricted or banned content [1]. However, the feature has functional limitations: it cannot access websites that require a login, nor can it bypass paywalls [1]. Additionally, the tool will not scrape URLs that have opted out via Google-Extended, the same crawler Google uses for training data [1].

The shift in AI strategy

The release of Deep Research coincides with a broader strategic pivot at Google regarding how it manages its AI research and product development. Historically, Google published hundreds of research papers annually, but the company has recently tightened its policy to prioritize incorporating findings into its own products before making them public [2]. This shift follows internal concerns that previous open-access research, such as the work behind the FLAN instruction-tuning technique, may have inadvertently aided competitors [2, 3].

Google is not alone in this transition toward agentic capabilities. Competitors are rapidly deploying similar features: Microsoft has introduced Copilot Vision for browser-based research, and Anthropic’s Claude now features tools that can control a cursor to execute computing tasks [1]. OpenAI is also reportedly developing its own agents for web browsing and coding [1]. As these companies move to keep their proprietary research closer to the vest, the focus has shifted from publishing academic breakthroughs to accelerating the speed at which AI models can execute real-world, multi-step tasks [1, 2].

What to watch

• Expansion of agentic features: Monitor whether Google integrates similar web-browsing capabilities directly into the Chrome browser, as reports suggest such a feature is currently in development [1].
• Competitive product releases: Watch for the rumored launch of OpenAI’s web-browsing and coding agents, which would directly challenge the utility of Google’s new research tool [1].
• Policy impact on innovation: Observe whether the restriction of public research papers by major firms like Google and OpenAI leads to a measurable slowdown in the pace of external, independent AI development [2].

The success of Deep Research will depend on its ability to navigate the tension between providing comprehensive, automated insights and respecting the growing number of websites that block AI scrapers. Whether this tool becomes a standard utility for knowledge workers or remains a niche feature depends on how effectively it can handle the increasingly restricted landscape of the open web.