deep research

basicwidespread

An AI mode in which an agent searches and reads on the web independently for minutes, then writes a source-backed summary report on your question.

In plain terms

A standard chatbot is like a fast-responding friend: it answers based on what comes to mind, perhaps quickly looking up one or two things. Deep research is more like a diligent intern you send to the library: it looks things up, takes notes, cross-references, and returns with an organized report.

In such cases, you do not receive an answer in seconds but typically between a few minutes and half an hour. Meanwhile, the agent reviews dozens, sometimes hundreds, of websites, studies, or uploaded documents, and finally delivers a multi-page report with references.

It is useful when there is no single good answer to your question, but the picture must be assembled from many sources: market overview, product comparison, background of a regulation, or literature review.

It differs from a search engine in that it does not provide links but a ready-made summary. It differs from simple chat in that it processes many more sources, methodically and in multiple rounds, and supports its statements with references so you can look them up.

How it works

You write the question; the system often asks follow-up questions to clarify what you are interested in and to what depth.
The agent creates a research plan: it breaks the big question into smaller sub-questions.
It searches sequentially, opens results, reads them, and initiates new searches based on what it read – this cycle is repeated many times, sometimes with multiple agents working in parallel.
It compares sources, tries to filter out repetitions and spot contradictions, then decides what goes into the report.
It writes the structured report and places sources next to the claims so you can verify them.

An everyday example

You want a heat pump for your family house but do not know which type is cost-effective. You enter the question into the chatbot's research mode, go for coffee, and twenty minutes later you find a multi-page comparison with prices, subsidy options, and links to sources.

Who is affected

employeeIt can reduce hours of background research, material gathering, and summarization to minutes, but you must verify the result. → Work & HR

parent, student, teacherIt provides a quick starting point for essays and theses, so schools need to reconsider what independent research means. → Education & Family

entrepreneur, executiveYou can get market and competitor analysis cheaply and quickly, for which you previously had to pay a consultant or analyst.

marketerAgents write about your company based on sources found on the web, so it matters what can be read about you and where. → Marketing

A common misconception

MythIf the report contains references, then every statement is verified and true.

In realityThe citation only shows where the agent obtained the information. It may still misinterpret the source, rely on a weak page, or make claims not present in the cited text.

Limits and risks

It can make mistakes and fabricate data just like any other language model; check important numbers and claims in the original source.
By default, it only sees the public web and what you upload or connect to it (for example, your own files, corporate storage). Due to paywalled articles, closed databases, and the scarcity of Hungarian-language sources, its knowledge may be incomplete.
It does not always distinguish reliable sources from ads or low-quality content, and a confident, polished report can hide this.
It is slower and more expensive than a standard response, so providers usually limit how many times you can use it per day or month.

Similar, but not the same

AI search – AI search provides short answers from a few results in seconds, while deep research works through many steps for minutes and produces a long report.
RAG – RAG is a technique: before writing a response, the model typically in a single step retrieves relevant text snippets and relies on them. Deep research goes further: it is a full, multi-round, independently planned research process.
scientific research – Deep research gathers and summarizes existing sources; it does not conduct new experiments, measurements, or discoveries.

Related concepts

Broader

AI-agent, generative AI

Narrower

research plan, source citation

On the edge

AI search, RAG, hallucination, multi-agent system, reasoning model

A brief history

Google introduced the Gemini Deep Research feature in December 2024, and OpenAI followed in February 2025 with the same name in ChatGPT. Throughout 2025, other major providers introduced their own versions, partly under different names (such as Research or DeepSearch), turning the product name into a general concept.

Where things stand

At the end of 2026, this is available in almost all major chatbots, in some cases even in the free tiers. Reliability and source handling have improved considerably, but human verification still cannot be skipped.

Status: 2026 November

Also known as

deep research · Deep Research · deep research · Deep research · Research mode · research agent · research agent · deep research agent

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.