ChatGPT Deep Research: How It Works and When to Use It

The workflow has four stages, and understanding them helps you write better queries.

1. Query Intake

You type a research question into the ChatGPT interface and select the Deep Research mode (it is a distinct option, not the default). You can also attach files: PDFs, spreadsheets, or documents that provide context to shape the research direction. The attached files are not published or shared; they guide o3's search strategy.

2. Research Planning

The o3 model reads your query and any attached files, then builds a research plan. This includes deciding which search terms to use, which types of sources to prioritize, and how to structure the final output. You can see a brief summary of the plan before the agent starts browsing, which gives you a chance to refine the scope.

3. Web Browsing and Source Collection

Deep Research executes its plan by running web searches, visiting pages, reading their content, and evaluating relevance. It is not limited to a single search. For a competitive intelligence query, for example, it might run 10 to 15 separate searches across different angles (pricing, features, market share, recent announcements) and visit dozens of pages per search. This stage is where the 5 to 30 minute time range comes from.

4. Synthesis and Report Generation

After collecting enough data, o3 synthesizes findings into a structured report. Each factual claim includes a numbered citation pointing back to the source URL. The report is delivered in your ChatGPT conversation as a formatted message, and you can export it or continue asking follow-up questions in the same thread.

Deep Research performs well on tasks where the value comes from aggregating information across many sources rather than generating creative output. Based on the feature's design and the types of queries it handles best:

• Multi-source literature reviews. Academic, technical, or industry research where you need to survey a broad landscape and identify consensus or disagreement across sources.
• Competitive intelligence. Comparing products, vendors, or market positions by pulling pricing, feature lists, reviews, and announcements from multiple sites.
• Market analyses. Industry sizing, trend identification, and landscape mapping that requires data from analyst reports, press releases, and vendor documentation.
• Technology landscape assessments. Evaluating tooling options, comparing open-source projects, or mapping integration ecosystems where no single source has the full picture.
• Due diligence research. Background research on companies, technologies, or regulatory environments where completeness and source diversity matter.

Deep Research is not the right tool for tasks where you already know the answer and just need it formatted, for creative writing, for code generation, or for conversational back-and-forth. Those tasks run faster and better on standard ChatGPT.

When your monthly Deep Research quota runs out, the feature does not disappear entirely. Instead, it drops into lightweight mode. This is an important distinction because the interface still shows Deep Research as available, but the behavior changes significantly.

In lightweight mode, Deep Research:

• Browses fewer web pages per query
• Consults fewer sources overall
• Produces shorter, less detailed reports
• May skip cross-referencing steps that full mode performs

The practical impact is that lightweight mode reports are closer in depth to what you would get from a single well-crafted Google search than the multi-source synthesis that full Deep Research provides. If you rely on Deep Research for work, plan your usage against the monthly cap. Running 10 exploratory queries on day one leaves you with lightweight mode for the rest of the month on the Plus plan.