In plain terms
An ordinary AI answer is like asking a well-read colleague in the corridor: the reply comes in seconds, from memory and a quick look. Deep research is like giving a junior analyst the afternoon. The analyst breaks the question into parts, looks things up in dozens of places, compares what turns up and comes back with a written report and a list of sources. The report still has to be read critically.
Why it matters
It compresses hours of desk research, such as a market scan, a supplier comparison or a regulatory overview, into minutes and returns a first draft with its sources. It suits broad questions that are well documented in public. Three limits decide how to use it. The report is only as good as the pages it found, and a vendor's blog can weigh as much as a regulator's notice. Wrong conclusions read as fluently as right ones. And a citation shows where a claim came from; whether the source supports it still has to be checked by a person.
Example
A strategy manager asks for an overview of the European market for home battery storage: size, main suppliers and subsidy schemes in five countries. The system asks two clarifying questions, then works for 18 minutes: 5 sub-questions, 60 searches, 140 pages read. It returns a 12-page report with 45 citations. She spot-checks ten of them. Two lead to pages that do not support the sentence, and one market figure is three years old.
Most often confused with
Deep Research vs. Answer engine
Both search the web and cite sources. An answer engine makes one pass, with a few queries and a short reply, which suits a factual question. Deep research runs as an agent: it plans, searches in rounds, changes direction according to what it finds and writes at length, which suits a question with many parts. It costs more and takes longer, and the length of the report makes errors harder to spot.
Origin: Google introduced a feature with this name in Gemini in December 2024; OpenAI and Perplexity followed with their own in February 2025.
Under the hood
The process is an agent loop on a reasoning model with tools for searching and reading pages. Typical stages: clarify the brief, write a research plan, split it into sub-questions, search and read in rounds (often with parallel subagents that each keep their own context), collect notes with their sources, then synthesise and attach citations, sometimes in a separate pass. Many versions also read uploaded files and connected company sources. A run usually takes from a few minutes to about half an hour and consumes many times the tokens of a chat answer, which is why subscription plans limit the number of runs. The name varies by product: Deep Research in ChatGPT, Gemini and Perplexity, Research in Claude, Researcher in Microsoft 365 Copilot. Weaknesses: uneven source quality, paywalled material it cannot read, outdated pages, citations that do not support the claim, and indirect prompt injection from the pages read. Evaluate with citation precision and with expert review of sampled reports.