Google adds Deep Research to Gemini

Background

Gemini is Google's AI assistant, and Deep Research was announced as a new capability inside its paid Advanced tier. The timing was deliberate: the same day, Google introduced Gemini 2.0 Flash Experimental, which it presented as the start of its "agentic era" — a way of describing AI systems that carry out a sequence of actions toward a goal rather than answering a single question^[2]. Deep Research was the first feature Google placed under that label for Gemini users.

What happened

Google announced Deep Research on 11 December 2024. The announcement was made by Dave Citron, Senior Director of Product Management for the Gemini app, in a post on Google's blog dated the same day^[1]. Google described the feature as a research assistant that does the legwork for the user — the phrase pointed at the tedious part of research: opening many pages, comparing them and keeping track of where each claim came from^[1].

How it works

Deep Research runs in stages. First it builds a multi-step research plan, which the user can revise or approve before any searching begins^[1]. Only then does it browse and analyse the web, a process that takes a few minutes rather than seconds^[1]. The output is a report, and the report carries links back to the original sources so that the reader can check the material behind each part^[1].

Google said the browsing and analysis are supported by a context window of 1 million tokens^[1]. A context window is the amount of text a model can hold in view at one time; a larger window lets the system keep many fetched pages in consideration together instead of discarding them between steps.

Access and details

At launch, users reached the feature by choosing "Gemini 1.5 Pro with Deep Research" from the model menu. It was available in English, on desktop and on mobile web^[1]. Reports could be exported to a Google Doc, so the result could be edited and shared like any other document^[1]. Google also said Deep Research would come to the Gemini mobile app and to Workspace accounts in early 2025^[1].

Why it matters

The design choices say something about how Google positioned the feature. The plan-approval step keeps the user in the loop before the model spends minutes working, in contrast with a single prompt and a single answer. The report with source links treats the output as something to be checked rather than simply read. And the export to Google Docs puts the result inside the tools where written work is normally finished.

Announcing it alongside Gemini 2.0 Flash Experimental, framed as the start of an agentic era, placed Deep Research as an early example of that idea in a consumer product: a system that plans, acts over several minutes, and hands back a finished artefact^[2]. The name also became a category — a research mode that spends minutes browsing on the user's behalf rather than answering immediately.