How do you use ChatGPT deep research for a report?

Deep research is a ChatGPT mode, launched on February 2, that does the legwork of a research assignment on its own. You give it a question; it searches the web, reads pages and PDFs, follows leads, sometimes backtracks, and returns a structured report with citations. It runs on a version of OpenAI's upcoming o3 reasoning model adapted for browsing and analysis, and a single task can take anywhere from 5 to 30 minutes.

For now it is available to ChatGPT Pro subscribers, with up to 100 tasks a month. OpenAI says Plus and Team users are next, and that it is not yet offered in the United Kingdom, Switzerland or the European Economic Area. It launched on the web, with the mobile and desktop apps promised within the month.

Because each run takes real time and counts against a monthly limit, the quality of your request matters more than in an ordinary chat. A vague question gets a long, vague report.

Step 1: Write a brief, not a question

Think of deep research as a capable research assistant you are briefing for the afternoon. Tell it four things: the question, who the report is for, the boundaries, and the shape of the result. Here is a worked example for a small business owner.

I run a three-person bakery in a small Canadian town and I'm considering offering online ordering with local delivery.

Question: What do small independent bakeries in Canada and the United States report about the costs, benefits and pitfalls of adding online ordering and delivery?

Audience: me, deciding within a month. I'm not technical.

Scope: focus on businesses under 10 employees; sources from 2022 onward; compare (a) third-party delivery apps, (b) an ordering page on my own website with my own delivery, and (c) pickup-only online ordering.

Format: a 2-page report with a summary table comparing the three options on startup cost, ongoing fees, effort and risks. Cite every number. Flag anything that is an estimate or comes from a company selling the service.

You can also attach files or spreadsheets for context, such as your current price list or sales figures, and the report will take them into account.

Step 2: Answer the clarifying questions

Before it starts searching, deep research often comes back with a few questions of its own. Early testers describe it asking for clarifications on scope before beginning; for the bakery brief, it might ask whether you mean one province or all of Canada, or whether you already have a website (the questions vary, so these are illustrative). Answer them briefly and specifically. This is the cheapest moment to steer the research: a one-line answer now saves a 30-minute run in the wrong direction.

Step 3: Let it work, and watch the sidebar

Once it starts, a sidebar shows a summary of the steps it is taking and the sources it is reading. You don't have to stay; ChatGPT sends a notification when the report is ready. Glancing at the sidebar is still worthwhile. If you see it leaning heavily on vendor marketing pages or drifting off topic, you will know which parts of the report need the most checking.

Step 4: Read the report as a draft

The finished report arrives in the chat as text, with citations attached to its claims. Read it once for structure: does it answer your question, cover all three options and include the table you asked for? Then read it for substance. OpenAI itself lists the main weaknesses: deep research can state facts that are wrong, draw incorrect conclusions, have trouble telling authoritative sources from rumor, and often fails to signal when it is unsure. It is also limited to what it can find and read during the run, so paywalled studies and many books are out of reach.

How do you check its sources?

Start with the claims your decision depends on, not with every sentence. In the bakery example, that means the fee percentages for delivery apps, any startup cost figures and anything presented as a typical result. Highlight five to ten such claims and audit only those. That is a manageable job and covers most of the risk.

For each one, click the citation and check three things. Does the page exist and load? Does it actually say what the report claims, with the same number, the same date and the same conditions? Testers have found that the links often jump straight to the passage being cited, which makes this faster. And is the source the right kind for the claim: an official fee schedule or a published survey, or a blog post, a forum comment or a vendor's sales page? A real link can still point to a weak source.

Keep a simple audit table as you go:

(The rows are illustrative.) Anything that fails the check comes out or gets confirmed elsewhere. If several key claims fail, ask deep research to redo that section using only official or independent sources, or do the remaining checking yourself. The report saves you hours of searching; the final judgment on what is true still belongs to you.