Using deep research in ChatGPT
Longer sourced investigations, research plans, source quality, and verification.
ChatGPT guide · as of July 10, 2026 · 4 minutes read · details change — confirm current specs on chatgpt.com
Deep research is for questions that cannot be answered well from one page or one quick search. You give ChatGPT a research objective, it searches and analyzes many sources, and it produces a structured report with citations.
The result can save hours. It can also be incomplete, overconfident, or built from sources you would not have chosen.
As of July 2026, access, limits, connected sources, and model behavior vary by plan and region. Confirm current details in OpenAI's deep research information.
When to use it
| Task | Quick search | Deep research |
|---|---|---|
| Current weather or score | Better | Excessive |
| Compare three products | Sometimes enough | Useful when evidence is scattered |
| Market or policy landscape | Too shallow | Strong fit |
| Literature orientation | Limited | Useful starting point |
| Find one official rule | Better | Often excessive |
| Build a sourced briefing | Limited | Strong fit |
Deep research is most valuable when the answer requires synthesis rather than one fact.
Write a research brief
Treat the prompt like a brief for a junior researcher.
Include:
- The decision or question.
- Audience and intended use.
- Geography and time period.
- Definitions and exclusions.
- Source priorities.
- Required comparisons.
- Output structure.
- What uncertainty must remain visible.
Example:
Research the current consumer market for private, local AI assistants in the United States. Focus on products ordinary households can install, not enterprise platforms. Compare hardware requirements, offline capability, voice support, privacy architecture, setup burden, and total first-year cost. Prefer official documentation, independent testing, and primary technical sources. Separate available products from announced products and speculation.
If the system asks clarifying questions, answer them. A vague brief produces a broad but less useful report.
Review the plan before the report
When possible, inspect the research plan or outline first. Correct missing categories before the system spends time gathering evidence.
Ask:
- Are the source types appropriate?
- Is the time window correct?
- Are important competitors or jurisdictions missing?
- Is the requested comparison actually measurable?
- Does the plan separate facts from forecasts?
The best time to correct the direction is before synthesis.
Audit the final report
| Check | Why it matters |
|---|---|
| Citation support | A linked page may not support the exact sentence |
| Source diversity | Ten articles may repeat one original report |
| Primary evidence | Marketing pages should not carry the entire conclusion |
| Date handling | Publication, event, and effective dates differ |
| Missing counterevidence | The report may converge too early |
| Geographic scope | US rules may be presented as global |
| Definitions | Sources may measure different things under one label |
Open the sources supporting the most important claims. Do not inspect only the easy or interesting ones.
Connected sources and private data
Deep research may be able to use connected apps or internal repositories in eligible accounts. This can combine company files with public web sources.
That is useful, but the data boundary becomes more complicated. Confirm which sources are active, whether the connection is read-only, and who can see the resulting report.
A document inside a connected system may contain malicious or irrelevant instructions. Treat it as content to analyze, not authority over the agent.
Do not connect a broad repository when a small approved folder would do.
Turn research into a durable artifact
A good final deliverable should include:
- Executive summary.
- Scope and definitions.
- Findings organized by the decision criteria.
- Comparison table.
- Evidence gaps and disagreements.
- Recommendations with conditions.
- Source list with access dates.
- A short maintenance note explaining what will become stale.
Save the report outside ChatGPT. Also save the prompt, source list, and date so someone can reproduce or update it.
What deep research is bad at
It can struggle with paywalled material, poorly indexed databases, local records, scanned documents, dynamic sites, and subjects where sources use inconsistent definitions.
It may cite secondary reporting when a primary document exists. It may miss a decisive source because search ranking did not surface it. It may also produce a polished recommendation where the evidence is genuinely mixed.
Ask it to include a section titled What this report could not establish.
When this is the wrong tool
Deep research is the wrong tool for emergencies, formal legal discovery, exhaustive systematic reviews, due diligence requiring licensed professionals, or any task where missing one source is unacceptable.
It is also unnecessary for a direct lookup that an official page answers clearly.
Use deep research to compress exploration. Do not outsource the final standard of evidence.
