Deep Research
Deep Research is an operating mode of AI assistants in which the program independently searches, reads, and summarizes many sources on the internet for a given question, compiling them into a longer report with citations. Instead of an answer in seconds, it delivers a document after several minutes that resembles a small piece of research work.
A normal chat with an AI assistant proceeds quickly. You ask a question, and a few seconds later the answer appears. Deep Research works differently. You give it a research task, and the program keeps working independently for several minutes. During this time, it searches the internet, opens dozens of pages, compares the information, and produces a coherent text from it. In the end, you don’t get a short paragraph, but a multi-page report with links to the pages used. The name is not a fixed technical term, but a product name that several providers use for their respective version of this feature.
From answer-giver to research assistant
Classic language models answer from what they absorbed during training. This knowledge has a cutoff date and is outdated on current topics. In addition, such models can invent things that sound plausible but are wrong. Deep Research tackles this problem by tying the answer to real, verifiable sources.
The second point is time savings. A proper market overview or a comparison of five providers can easily cost a person several hours. The system handles the collection work in a few minutes. That is exactly why the feature is often described in business news as an attack on professions that live off gathering information: market research, consulting, parts of journalism.
An important distinction concerns what lies below it. An ordinary web search in a chat fetches two or three hits and summarizes them. Deep Research plans the research over many steps and decides along the way where it still needs to dig deeper. So the difference is not just the length of the result, but the independence of the approach.
Search, read, follow up: the process
At the start, the model breaks the task down into sub-questions. From “What is the state of the European battery market?” come questions about manufacturers, capacities, subsidy programs, and prices. For each sub-question, the system formulates its own search queries. Many systems first ask clarifying questions to sharpen the task.
Then a loop of searching, opening, and reading begins. The model calls up pages, extracts the relevant passages, and stores them in a temporary buffer. If two sources contradict each other, it can specifically search for a third. This loop often runs fifty times or more until enough material has been gathered. Only afterward does the model write the report and attach evidence to the individual statements.
A common misconception is that the citations replace verification. The model links what it has read, but assesses credibility only to a limited extent. A forum post and an official statistics agency can end up side by side. The attribution is also not always correct: sometimes a source is attached to a sentence that doesn’t actually support it. Spot-checking therefore remains mandatory.
Who offers the feature and what it costs
OpenAI, Google, and Anthropic each have their own Deep Research features in their chat interfaces. You usually find them as a toggle or menu item next to the input field. After starting, the task runs in the background, and you are notified when the report is ready. Perplexity and a few smaller providers have comparable modes.
Because each research task consumes a lot of computing time, usage is limited almost everywhere. In free accounts, it’s often only a few reports per month; in paid subscriptions, a few dozen. This is a recurring topic in news coverage: providers advertise the feature but cap it, because the cost per task is significantly higher than that of a normal chat.
In everyday use, Deep Research pays off wherever a question touches many scattered sources: a term paper on a current topic, comparing degree programs, an overview of a company before a job interview. For simple factual questions, the feature is overkill. You then wait five minutes for something that an ordinary search delivers in seconds.