Deep Research

Deep Research refers both to a category of agentic AI behavior and to specific product features in which a language model operates as a semi-autonomous research assistant. The agent decomposes a high-level research goal into sub-questions, executes iterative web searches and document reads, evaluates source relevance and credibility, follows promising links, and updates an internal knowledge state across many steps before producing a structured, cited report. The defining characteristic is that the agent determines its own research trajectory rather than responding to a single-turn query.

The workflow typically begins with a planning phase in which a reasoning-capable model outlines the information needed to address the query. It then enters a loop of searching, reading page content, extracting relevant facts, and deciding whether to pursue new leads or consolidate findings. Sessions commonly involve dozens to over a hundred web requests and run for several minutes to half an hour on complex topics. The final output is a long-form report with inline citations, often structured with headings, tables, and summaries.

OpenAI released a product feature named Deep Research in February 2025, built on the o3 reasoning model; Google launched a feature under the same name within Gemini in late 2024 and expanded it through 2025. Perplexity offered a comparable mode earlier. These tools found rapid adoption for competitive intelligence, academic literature surveys, and due-diligence reporting—tasks that previously required hours of manual effort from knowledge workers.

As of 2026, Deep Research is a standard capability tier offered by major AI platforms, and enterprise deployments routinely augment web-browsing access with retrieval over internal document stores. Persistent challenges include hallucination of sources not actually read, over-reliance on SEO-optimized content, and limited access to paywalled or proprietary databases. Evaluation frameworks for research agent output quality—covering citation accuracy, coverage, and factual correctness—have become an active area of benchmarking.

Example

“A market analyst prompts a Deep Research agent with a question about the competitive landscape of vector database vendors, and the agent autonomously searches industry reports, vendor documentation, and technology news over roughly 20 minutes, returning a structured report with a comparison table and numbered citations.