Launch a Deep Research Task
On this page
Your Initial Request Determines Everything
The quality of a Deep Research report depends directly on the quality of your request. This is a hard rule with no exceptions: a well-crafted prompt yields a precise, actionable report, while a vague prompt produces a superficial overview you won’t dare show anyone. You cannot fix this gap after the fact because the research plan is built entirely on what you wrote. Spend three minutes refining your request before launching the twenty minutes of processing that follow.
Accessing the Feature
OpenAI’s help center lists three ways to launch a deep research task:
- Type /Deepresearch directly into the input field
- Select Deep research from the tools menu (the + button)
- Choose Deep research from the sidebar
Before writing your prompt, select your sources if needed. By default, the search can use the public web, files you upload, and compatible connected apps. To focus the search on specific sites, go to Sites > Manage sites: you can limit the search to listed domains or prioritize them while still including the rest of the web.
ChatGPT may then ask clarification questions and propose a research plan. Review and edit this plan before launching the task. Answer any clarification questions precisely; these preliminary exchanges significantly improve the final report. Dismissing them with “whatever” lets the tool define the scope for you, and you will likely find the delivered report misses your actual needs. The proposed plan is your final checkpoint: tweak it rather than accepting it by reflex.
The Four Components of a Good Prompt
An effective Deep Research prompt rests on four pillars. The context states who you are and why you need this information; this filters out anything irrelevant to your situation. The central question specifies exactly what you want to know, preventing the research from scattering. The scope sets geographic, temporal, and sectoral limits, often providing the highest return on effort. The desired format dictates how the report should be structured, saving you from reformatting everything later.
Compare these two formulations to see the difference. The basic version—“Tell me about electric cars in France”—yields poor results because it lacks all four pillars. The optimized version includes them all:
“I am a fleet manager for a company with 200 employees in the Île-de-France region. I need to decide whether to switch our fleet of 50 vehicles to electric in 2026. Provide a comprehensive analysis covering: total cost of ownership (purchase, maintenance, energy) compared to internal combustion engines, available grants and subsidies in 2026, the state of the charging network in Île-de-France, and feedback from similar companies. Structure the report with an executive summary at the beginning.”
The report generated from this second prompt is ready to present to a committee because the request already defined its purpose.
Templates to Adapt, Not Copy
For a market study, a proven structure is: “Analyze the [sector] market in [geographic area] in 2026. Include: market size, trends, key players, opportunities, and threats. Cite your sources.” This works because it enumerates the expected sections, ensuring none are missed.
For a competitive benchmark, ask: “Compare [Company A], [Company B], and [Company C] on the following criteria: [criterion 1], [criterion 2], [criterion 3]. Present the results in a comparison table with sources.” Naming the criteria prevents you from receiving a comparison on axes that don’t interest you.
For regulatory monitoring, try: “What new [sector] regulations have entered into force or are planned in [country] for 2026–2027? Detail the impact for a company with [your profile].” Mentioning the profile transforms a list of texts into an impact analysis.
These templates are starting points, not fill-in-the-blank forms. A copied prompt that isn’t personalized with your sector, region, and profile will produce a generic report anyone could have obtained.
While the Research Runs
You can track progress in real time and interrupt the search at any moment to refocus it or modify allowed sources. However, avoid launching multiple tasks simultaneously: your quota is limited, and so is the attention you can devote to each report.
Use this time to observe the selected sources and compare them to what you would have found on your own.
Your Turn
Your first deep research task. Launch a trial run with a prompt like this, adapted to your context, and review the proposed plan before validating it:
“I work in [your sector]. Analyze the 5 most impactful trends in my sector for 2026. For each, include: a description, key figures, concrete examples of companies adapting to it, and implications for a company of [your size]. Structure the report with an executive summary at the start.”
Official sources: Deep research in ChatGPT (OpenAI Help Center)
Key Takeaways
- Launch Deep Research via /Deepresearch, the + button, or the sidebar
- A good prompt includes: context, central question, scope, and desired format
- Answer clarification questions to improve results
- Review and edit the proposed plan, then track progress and interrupt if needed
- Choose your sources: public web, priority or exclusive sites, files, or connected apps