The short answer.

Deep research is a mode in ChatGPT, Claude, Copilot and Gemini where the assistant works on one question for several minutes: it searches, reads many sources, follows what it finds and returns a structured report with its sources cited. Use it for market scans, vendor comparisons and background before a meeting, where a quick answer would be too thin. Treat the report as a well-organised first draft: it reads weak sources as readily as strong ones, and a person checks the key citations before anyone acts on it.

What each maker calls deep research, and where it reads from.

Each of the four makers has its own name and its own mix of sources for deep research. Makers rename and move these features often; what follows was read on their own help pages in November 2026.

  1. ChatGPT.OpenAI calls it deep research. ChatGPT proposes a research plan you can change before it starts, shows its progress, and lets you interrupt to narrow the focus. It reads the public web and files you upload, and it can read from apps your company has connected, such as a document store, without writing anything back. You can tell it to stay on a list of websites, or to prefer them.
  2. Claude.Anthropic calls it Research. Claude runs a series of searches that build on each other and decides what to look into next from what it finds. Web search has to be on. It can also search the connectors you have turned on, such as your Google Workspace mail, calendar and documents. Started from the phone app, it searches the web only.
  3. Copilot.Microsoft calls it Researcher, and lists it among Copilot’s agents. Researcher may ask you follow-up questions first, then returns a report with key insights, cited sources and next steps. At work it reads the web and the files, emails, meetings and chats you already have access to, under your company’s existing permissions. Your administrator decides whether it is switched on.
  4. Gemini.Google calls it Deep Research. Gemini drafts a research plan you can edit, then runs while you do something else and tells you when the report is ready. It searches Google by default, and can add Gmail and Drive if your Google Workspace is connected, files you upload and notebooks from NotebookLM. The finished report exports to Google Docs.

The biggest difference between them is what each one reads besides the web. An assistant that can read your own mail, files and meetings writes a different report from one that only reads the internet. Google also offers deep research to developers through its programming interface, for building into software; that lives in a developer tool, not in the everyday Gemini app.

The work deep research suits.

Deep research suits a question that needs many sources read and compared, where the answer is a page or more of writing, not a fact. Three kinds of work fit it best in a company of 10 to 200 people.

  • Market scans.Who sells into a segment you are thinking of entering, how they describe themselves, what has changed in the last year.
  • Vendor comparisons.Four accounting systems, five field-service tools or three payroll providers set side by side on the points you care about, from their own pages and public reviews.
  • Background before a meeting.A prospect, a supplier or a partner: what they do, what they have announced, who leads them, what they are likely to raise.

Deep research is the wrong tool for a quick fact, a calculation or a decision that turns on your own numbers. A plain chat with web search is faster for the first, and a person with the spreadsheet is right for the others. What work should you not give to AI? covers where that line sits.

The advantages of deep research.

The main advantage of deep research is breadth in one sitting: the assistant reads more sources than a person would open for a routine question, and lays them out in one structured report.

  • It shows its plan.ChatGPT and Gemini show you the research plan before they start, and Copilot asks follow-up questions. You correct the question before the work runs, not after.
  • It cites its sources.Each claim can be traced back to a page, which makes the report checkable in a way an ordinary chat answer often is not.
  • It can read your own records.Where your company has connected mail, files or documents, the report can set what the web says beside what your own records say.

The trade-offs of deep research.

The trade-offs of deep research are time, source quality and the checking a person still has to do.

  • Time.Deep research takes several minutes where a chat takes seconds, and each maker limits how often you can run it. Save it for questions worth the wait.
  • Source quality.Deep research reads what it can reach. A vendor’s own marketing, an old forum thread and a trade body’s figures can all sit in one report with equal weight. Sources behind a paywall or a login are often out of reach, so the best source on a subject is often missing.
  • A citation is not a confirmation.A cited page may say something narrower than the sentence that cites it, or be out of date. The report reads with confidence either way.
  • Connected records widen what it sees.A report that reads your mail and files can surface material the reader of the report should not see. Decide who runs research across company records before anyone does.

None of these is a reason to avoid deep research. Each is a reason for a person to read the report as a draft and check it before it reaches a decision. How to check work that AI prepared covers that habit for every kind of AI output.

A checklist for reviewing a research report.

Copy this list and run it on any deep research report before it goes into a decision, a proposal or a meeting.

  1. Was the question right?Read the plan or the opening of the report. Check it answered your question, not a broader or easier one.
  2. Open the three claims that matter most.Follow each citation to the page. Check the page says what the report says, in the same scope.
  3. Who wrote each key source?Mark each as the company itself, an independent source or an anonymous one. Weigh them accordingly.
  4. How recent is it?Check the dates on the sources behind any claim about prices, features, people or market position.
  5. What is missing?Name the source you would have checked first. If the report never read it, read it yourself.
  6. Which sentences have no citation?Treat every uncited conclusion as the assistant’s opinion.
  7. What did it read from our own records?Where connected mail or files were used, check nothing private is passed on to people who should not see it.
  8. Who signs it off?One person, by name, reads the checked report and decides what goes forward.

One brokerage prepares for a meeting.

Take a 30-person insurance brokerage whose principal is meeting a carrier the firm has never placed business with.

Illustrative example.

The account executive runs deep research on the carrier: what it writes, what it has announced this year and how it describes its appetite for the firm’s kind of client. The assistant’s plan includes the carrier’s financial history; she removes that part, because the firm already uses a rating service for it, and adds the carrier’s published underwriting guidelines. She opens the four that carry the main points, finds one is a news piece from two years ago describing an appetite the carrier has since changed, strikes that paragraph and sends the principal a one-page brief with the checked sources listed. The principal decides what to propose.

Getting more from the AI you already pay for.

Deep research is one of many features in the assistants your company already pays for that most people never open. FullyPowered AI helps companies level up what they get from that AI: we plan its use with you, develop the workflows and deploy them to the people who do the work. The plan lives in FullyPowered Planner, and How organizations plan their use of AI sets out the method.