Deep research (AI research mode)

What is deep research?

Deep research is a mode in the big AI assistants where you hand over a question and walk away for a while. Instead of answering from what the model already knows, an agent writes a research plan, runs dozens of web searches, opens and reads the pages it finds, searches again based on what it read, and after several minutes returns a long report with a citation behind most claims.

OpenAI put the name on the map when it shipped deep research in ChatGPT in February 2025, describing it as an agent that will "find, analyze, and synthesize hundreds of online sources to create a comprehensive report" and warning that a run "may take anywhere from 5 to 30 minutes". Google had launched Deep Research in Gemini two months earlier, Perplexity followed within two weeks with a version that finishes in two to four minutes, Anthropic added Research to Claude in April 2025, and Microsoft built its Researcher agent in Microsoft 365 Copilot on top of OpenAI's deep research model. The names differ slightly, the shape is the same everywhere.

Think of it as sending a junior analyst to the library for an afternoon. You give them the question, they come back with a memo and a reading list. The memo is usually good. The reading list is where you find out whether they actually read the books.

How a deep research run works

Anthropic's engineering team published the clearest description of what happens under the hood when it wrote up how it built Claude's Research feature. The pattern is called orchestrator-worker, and most deep research products work along the same lines.

A lead agent plans. The lead agent reads your question, thinks through an approach, and writes that plan down before it does anything else. Several products show you this plan first: Gemini lets you edit the plan before you press start, ChatGPT and Copilot's Researcher often ask a clarifying question or two. That minute of steering is worth taking, because it is the last moment you can cheaply change direction.

Subagents search in parallel. The lead agent hands out pieces of the question to subagents, each with its own fresh context window, and each one runs its own searches and page reads. Anthropic's guidance for its own system gives a feel for the scale: a simple fact-finding question gets one agent with 3 to 10 tool calls, a direct comparison gets 2 to 4 subagents with 10 to 15 calls each, and a complex research question can use more than 10 subagents.

The search is iterative. Good runs start with short, broad queries, look at what comes back, then narrow down, the way a person would. The lead agent collects what the subagents found, decides whether it has enough, and if not sends out another round. This is why the process cannot be a fixed pipeline: nobody knows in advance which page will open the next lead.

A separate step adds citations. In Anthropic's system a dedicated citation agent takes the finished report and the documents that were read, and works out where each citation belongs. Citations are added after the writing, not during it. Keep that in mind when you read the next section.

All of this costs tokens. Anthropic measured that an agent uses roughly four times the tokens of a normal chat turn, and a multi-agent research run about fifteen times. That is why every vendor caps the number of runs: after an April 2025 change, ChatGPT Plus users get 25 full deep research runs a month and free users 5, and Gemini and Perplexity apply daily limits that rise with the paid tier.

Deep research versus a chat answer, and versus RAG on your own documents

Versus a normal chat answer: how much work happens before the first word

A chat answer is one model call. The model may run a quick web search, but it answers in seconds from its training data plus whatever the top results said. Deep research spends minutes and dozens of tool calls before it writes a word, and the report is built up from what it read rather than from what it remembered. Ask a chat model for the current price of a competitor's product and you get a plausible number. Ask deep research and you get a number with a page behind it, which you can open.

Versus a RAG assistant: where the sources come from

A RAG assistant over your own documents and a deep research agent look alike from the outside, both return an answer with references. The difference that matters is where the sources come from. RAG searches a set of documents you chose: your contracts, your manuals, your SharePoint. It cannot know anything outside that set, and that is its strength: when it cites a source, the source is yours and you can hold it to it. Deep research searches the open web, which nobody curated. It can find things you did not know existed, and it can also find a forum post from 2019 and a press release from a competitor and weigh them the same.

The vendors are closing the gap from both sides. Gemini Deep Research can include your Gmail and Drive as sources, Claude's Research reads Google Workspace alongside the web, and Microsoft's Researcher combines your files, mails and meetings with web results in one report. That is useful, and it makes the question sharper rather than moot: for every claim in the report, which side did it come from?

What deep research is good for in a small company

Deep research earns its place on questions where the answer is spread across many public pages and nobody in the company has an afternoon to read them all.

1. A market scan
   "Which companies in Belgium and the Netherlands sell industrial label printers to the food sector, and how do they position themselves?" The agent will find a list you would not have assembled by hand, and it will miss the two firms that only exist on LinkedIn.

2. A supplier check
   Data from the company register (the KBO in Belgium), filed annual accounts, news, court records, reviews. A good first pass before you sign, provided you then open the register page yourself.

3. A regulation summary
   "What does the EU packaging regulation mean for a company that ships 200 parcels a day, and by when?" Deep research is strong at pulling the relevant articles and dates together. It is weaker at knowing which delegated act was published last week.

4. Competitor pricing
   Public price lists, webshop pages, tender awards. Ask for the pricing table with a source URL per row and you get something you can maintain.

What to watch out for with deep research

It only reads what it can reach. Paywalled journals, trade databases, sites that block bots and PDFs that need a login are invisible to the agent. So is anything published after the search index caught up. A report on a topic where the good sources sit behind a paywall will be built from the free sources that discuss them, which is a different thing.

The citation may not say what the sentence claims. Because citations are attached after the writing, a sentence can be a reasonable synthesis of three pages and end up with one citation that supports only part of it. A study of ten models and agents published in April 2026 by a team at the University of Pennsylvania found that between 3 and 13 percent of citation URLs in these reports point to pages that never existed, and that deep research agents cite more per query than search-augmented chat models but fabricate URLs at higher rates. A citation is a pointer, not proof.

Thin evidence reads as confident prose. OpenAI's own launch page says the model "may struggle with distinguishing authoritative information from rumors" and "shows weakness in confidence calibration, often failing to convey uncertainty accurately". In practice: two blog posts repeating each other come out as an industry consensus. The report does not get less fluent when the evidence gets thinner.

The pages it reads can talk back. OpenAI's system card names prompt injection as a specific risk: if a page the agent reads contains instructions, the model may follow them and return an inaccurate answer, for instance recommending the wrong product. On a supplier check, that means the supplier's own page has a say in the report.

Two runs, two reports. The search path is dynamic, so the same question on Monday and Tuesday can open different pages and reach different conclusions. That is not a bug you can turn off; it is a property you can use, see below.

How to check a deep research report

You will not verify a 4,000-word report line by line, and you do not have to. Four habits catch most of the damage.

Spot-check three citations. Pick the three claims you would act on, open their sources, and read the passage itself, not just the page title. If one of the three does not say what the report says, treat the whole report as a draft and check the rest of the claims you care about.

Read the search log. ChatGPT shows a sidebar with the steps taken and the sources used, Gemini shows its reasoning as it goes, and the others keep something similar. Skim it for what the agent did not find: the search it gave up on, the site it could not open. The gaps tell you which part of the report is synthesis rather than reading.

Run it twice. Ask the same question in a fresh session, ideally in a different product. Where the two reports agree, the evidence is probably solid. Where they disagree, you have found the exact spot that needs a human.

Ask what it could not find. End the prompt with a request to list the questions it could not answer from the sources it reached. A report that admits three gaps is more useful than one that fills them.