Workbranch › Guides

How to use ChatGPT deep research

Deep research sends ChatGPT off to read many sources and come back with a cited report. The quality depends mostly on your brief and on how carefully you check the result. Here's how to do both.

Updated 18 October 2026

These steps were checked against OpenAI's help pages on 18 October 2026. Menu labels may move as the interface changes.

When to use it (and when not to)

Deep research is for questions that need many sources combined and weighed: a market overview, a comparison of approaches, the background for a decision. It plans its own searches, reads, refines and writes up what it found. For a quick fact, a recent headline or a single number, normal ChatGPT search is the better tool: it usually answers in a few seconds.

Step 1: Start a task

1. Open a new chat in ChatGPT.
2. Start deep research in one of three ways: select Deep research from the + tools menu in the message box, select Deep research in the sidebar, or type /Deepresearch in the prompt.
3. Attach any files you want it to use, using the same + menu. See how to upload files to ChatGPT for formats and limits.

Step 2: Choose the sources

By default deep research can use the public web, files you upload and supported connected apps. To steer it toward particular websites:

1. In the message box, open Sites and then Manage sites.
2. Enter the sites or domains, one by one or as a comma-separated list.
3. Choose to search only those sites, or pick Prioritize these sites, but allow full-web search to favor them while still looking wider.

Restricting sources is useful when you trust a few publishers, such as a regulator, a standards body or your own company's documentation, and don't want random blogs mixed in.

Step 3: Write a good research brief

OpenAI recommends stating the question, the outcome you want and the constraints. A brief that covers these five points rarely needs rework:

• The question. One sentence on what you need to know.
• Who it's for and why. A board decision, a client meeting, your own learning. This sets the depth and tone.
• Scope. Timeframe, region, industry, and anything to leave out.
• Sources. Which kinds of sources to trust or avoid, and whether to use your attached files.
• Format. Length, sections, a comparison table, a list of recommendations.

An example: "I'm deciding whether our 20-person design studio in Portugal should switch to a four-day week. Research what's known from trials and studies since 2020, focusing on small creative businesses in Europe. Prefer government reports and peer-reviewed studies. Give me a two-page report with the main findings, the risks, open questions, and three practical options, each with its evidence."

If ChatGPT comes back with clarifying questions, answer them; that's part of the brief. Then read the proposed research plan and edit it before it starts. Fixing the plan is far cheaper than fixing the report. For more on prompts that lead to decisions, see prompts for actionable answers.

Step 4: Follow it and adjust

You can watch progress while it runs. If it heads in the wrong direction, interrupt it to change the focus or the sources it may use. OpenAI says a run may take about 5 to 30 minutes, and you'll be notified when the report is ready.

Step 5: Review the report and its sources

The finished report opens full screen with a table of contents, a sources used section and an activity history that shows how the research went.

1. Read the summary, then open the citations behind the claims that matter most to your decision.
2. Check that each source says what the report says it does, and that it's recent and from someone credible.
3. Use the activity history to see which searches it ran and what it may have missed.
4. Ask follow-up questions in the same chat to fill gaps or go deeper on one section.
5. Download the report as Markdown, Word or PDF to share it.

Treat the report as a well-researched draft, not a verdict. Citations make checking easy, but you still have to check, especially before anything goes to a client or into print.

Good to know

• Limits vary by plan. The in-product counter shows your remaining tasks. On plans with a fixed monthly allowance, it resets every 30 days from your first use.
• Models. It uses the latest model by default, and you can pick a legacy model for a task if you prefer.
• Apps are read-only. Deep research only uses read actions from connected apps, and only apps that support it. In Enterprise and Edu workspaces, admins control who gets access.
• Privacy. Deep research chats follow your normal data controls and chat retention.

If it doesn't work

• No Deep research option. It may not be offered on your plan or in your country, or your workspace admin hasn't given your role access.
• A connected app isn't offered as a source. Not every app supports deep research. Check the app is connected to an account that can open the files, and that your workspace allows it.
• You've run out. Check the usage counter for when it resets, and use regular search in the meantime.
• The report is off target. Tighten the brief, restrict the sites and edit the plan before starting again.

Turning the report into work

A long report often ends in a list of things to do. Workbranch, a separate AI workspace whose answers come from Google Gemini, can hold that follow-up as a point-by-point checklist with progress, due dates and reminders. It has a 3-day free trial. Our post on turning AI answers into tasks covers the approach.

Sources

• Deep research in ChatGPT (OpenAI Help Center, checked 18 October 2026)
• Research with ChatGPT (OpenAI Academy, checked 18 October 2026)

Frequently asked questions