Gemini Deep Research

What it is

Gemini Deep Research is a feature inside the Gemini app that does research on the user's behalf instead of answering in a single reply. Google launched it on 11 December 2024 for Gemini Advanced subscribers, on desktop and on the mobile web, starting in English ^[1].

How it works

The feature is built around a plan the user sees before any browsing happens. Given a question, it first drafts a multi-step research plan, which the user can revise or approve as it stands. Only then does the tool browse the web, spending a few minutes on the task, and it returns a report whose points are linked back to the original sources it read ^[1].

That order matters. A single search query is a guess about where an answer lives, while a research plan breaks the question into sub-questions that can be searched one by one and then put back together. Letting the user edit the plan first creates a cheap place to catch a misreading of the question, before minutes of browsing are spent on the wrong thing. The links in the finished report serve a similar purpose: a reader who doubts a sentence can follow the citation back to the page it came from and check it.

At launch the feature ran on Gemini 1.5 Pro ^[1]. A finished report can be exported to Google Docs ^[1], so it can be edited further or passed on to someone else without copying text by hand.

Where it sits

Google described the feature as agentic research: the system is given a goal, works out intermediate steps, carries them out, and comes back with an artefact rather than a chat reply. At launch, Google said Deep Research would reach the Gemini mobile app and Workspace accounts in early 2025 ^[1].

The capability later became something developers could call directly. Google offered Deep Research to developers as a managed agent in the Gemini API: deep-research-pro-preview-12-2025 was released on 11 December 2025, and was followed on 21 April 2026 by deep-research-preview-04-2026 and deep-research-max-preview-04-2026 ^[2]. A managed agent in this sense means the planning, browsing and report-writing loop runs on Google's side; an application that uses it sends the question and receives the result, rather than building and maintaining that loop itself.

Taken together, the two steps describe the same capability in two forms: a feature that a person opens in an app and works through with a plan in front of them, and a service that another program calls in the background. The parts that stay constant are the shape of the output — a multi-step plan, a wait of some minutes, and a report tied to the pages it came from.