How to use ChatGPT deep research: step by step, with a worked example and 5 prompts

Deep research turns ChatGPT from a quick-answer tool into a research assistant that reads many sources and writes a cited report. Here is how to brief it well, steer it and check what it finds.

The short answer

ChatGPT deep research is OpenAI's mode for multi-source questions: it plans the research, reads many sources and returns a report with citations. Start it by typing /Deepresearch, choosing Deep research from the + tools menu or picking it in the sidebar, then choose sources, review the proposed plan and let it run. Reports download as Markdown, Word or PDF. Deep research is on every plan, with a limited allowance on Free and the most on Pro.

What ChatGPT deep research is

Deep research is the ChatGPT mode for questions that need many sources, not one quick answer. Instead of replying in seconds, ChatGPT plans the research, reads and compares a large number of sources, and hands back a written report in which every finding links to where it came from. OpenAI's help centre suggests ordinary search for quick facts and deep research when you need depth and thoroughness: a market scan, a competitor review, a policy comparison, background before a big meeting.

It is on every ChatGPT plan with different allowances, although OpenAI says availability also depends on your country. OpenAI's pricing page lists a limited allowance on Free, expanded deep research on Plus and the maximum on Pro. The general idea is explained in what is deep research.

What you need first

  • A ChatGPT account on any plan. Usage varies by plan and is limited on most, so save it for questions that need it. An in-product usage counter shows the tasks you have left, and on plans with a fixed monthly allowance it resets every 30 days from your first use.
  • A clear research question, with the decision it supports. "Everything about the EU AI Act" is too wide; "what the EU AI Act requires of a 50-person company that uses AI for hiring, and by when" is a research brief.
  • Your own files, if relevant, such as a previous report or data export to upload.
  • A list of trusted websites, if you want the research limited to official sources.
  • Time. A deep research run takes minutes, not seconds, so start it and get on with something else.

How to use ChatGPT deep research step by step

  1. Start deep research. According to OpenAI's help article, you can type /Deepresearch in the message box, choose Deep research from the tools menu (the + button), or select it from the sidebar.
  2. Choose the sources. By default it can use the public web, files you upload and connected apps that support deep research, depending on your plan, region and workspace settings. To focus on specific websites, open Sites, then Manage sites in the prompt window, and either restrict the research to those sites or choose Prioritize these sites, but allow full-web search. Deep research only uses read actions in connected apps.
  3. Write your research brief. Give the question, the scope (countries, dates, industries), the sources to prefer and the shape of the report. The prompts below are ready to adapt.
  4. Answer its clarifying questions and review the plan. ChatGPT may ask questions and propose a research plan before it starts. Edit the plan if it has misread the scope. This is the cheapest moment to fix a bad brief.
  5. Watch progress and steer. You can follow the work as it happens, interrupt to change the focus and update which sources it may use.
  6. Read the report with the sources open. Every deep research output includes citations or source links. Click the ones behind the findings you will act on.
  7. Download or keep working. The finished report opens in a fullscreen view with a table of contents, a sources used section and an activity history. Reports can be downloaded as Markdown, Word or PDF. You can also ask follow-up questions in the same chat, such as "turn section 3 into a one-page brief for the leadership team".

Steps follow OpenAI's help article, Deep research in ChatGPT. OpenAI changes ChatGPT often, so if a menu looks different in your app, the article is the place to check.

A worked example: a supplier market scan

An operations manager needs to shortlist cloud backup providers for a 40-person firm in Germany before a budget meeting on Friday.

  1. She starts deep research and limits the sources to the providers' own websites plus Germany's Federal Office for Information Security, with the wider web as a lower priority.
  2. Her brief: "Compare five business cloud backup providers for a 40-person company in Germany. For each: where data is stored, encryption, certifications, restore options, published pricing model and contract length. Use official pages; mark anything you cannot confirm. End with a table and a recommended shortlist of two with reasons."
  3. ChatGPT asks whether she needs on-premise backup too. She says no, and edits the plan to drop one provider that does not serve Europe.
  4. The report arrives with a table and links. She opens the pricing and data-location sources for the two shortlisted providers, finds one claim based on an old page, and corrects it.
  5. She downloads the report as Word and attaches it to the budget paper.

An afternoon of tab-switching became twenty minutes of briefing and checking. The checking is not optional: see how to check if AI is telling the truth.

Common mistakes with deep research and how to fix them

  • A vague question. Deep research will dutifully research everything. Fix: add the decision, the scope and the output format.
  • Skipping the plan. Fix: read the proposed plan and edit it before it starts.
  • Using it for quick facts. It wastes an allowance that is limited on most plans. Fix: use normal search for one-line answers.
  • Trusting every citation. A link only shows where a claim came from, not that the claim is right. Fix: open the sources for the claims you will repeat.
  • Letting it use any website. Fix: restrict or prioritise official and primary sources.
  • Uploading confidential files to a personal account. Personal Free, Plus and Pro workspaces can be used for training unless you switch off Improve the model for everyone under Settings, then Data controls. Business and Enterprise are not used for training by default.

ChatGPT deep research or another tool?

  • ChatGPT is the easiest default because deep research is on every plan, including a limited allowance on Free.
  • Gemini Deep Research is also free. You start it from Add Files, then Deep Research, can include Gmail and Drive if Google Workspace is connected, click Edit plan before Start research, and export the report with Export to Docs. Google says reports usually take 5 to 10 minutes. Better if you live in Google Workspace; see how to use Gemini.
  • Claude Research needs a paid plan (Pro and above). Click +, then Research, or type /deep-research in the new experience. It runs on Claude Opus 5.5. See how to use Claude.
  • Perplexity lists Deep Research from the Pro plan, and every ordinary answer already cites its sources. See how to use Perplexity.

For a full comparison, read best AI for research, ChatGPT vs Perplexity and our prompts for research. Our free ChatGPT lessons cover deep research with exercises.

5 deep research prompts to copy

Each one is a full research brief. Replace the [placeholders].

Prompts to copy

Market scan with a shortlist

Research [number] providers of [product or service] for a [size] company in [country]. For each, cover [criteria 1], [criteria 2], [criteria 3] and the published pricing model. Use official company pages first. Mark anything you cannot confirm as unconfirmed. End with a comparison table and a shortlist of two with reasons.

You get: A sourced comparison table and a shortlist

What a new rule means for us

Explain what [law or regulation] requires of a [type and size of organisation] in [country], what the deadlines are and what changed in the last 12 months. Prioritise official government and regulator sources. Finish with a checklist of actions in order of deadline, each linked to its source.

You get: A compliance brief with a dated action list

Competitor review

Research [competitor A], [competitor B] and [competitor C]: their product lines, published pricing, recent launches in the last six months and how they describe their target customer. Use their websites, release notes and reputable press. Then compare them with us: [paste our positioning]. Give three opportunities and three threats, each with evidence.

You get: A competitor brief with evidence for each point

Background before a meeting

I am meeting [person, role] at [organisation] on [date] about [topic]. Research the organisation's strategy, recent news from the last 12 months and anything public the person has said about [topic]. Keep it to two pages with links, and end with five smart questions I could ask.

You get: A two-page briefing and questions to ask

Evidence review

What does the recent evidence say about [question]? Use studies, official statistics and reviews from the last [three] years. For each key finding, give the source, the year, the sample or scope and how strong the evidence is. Point out where studies disagree and what is still unknown.

You get: A balanced evidence summary with its limits

Start the ChatGPT lessons Start the Gemini lessons