Deep research mode

What is deep research mode?

Deep research mode is the agentic, multi-step research feature in AI search and assistant products that autonomously runs many web searches, reasons across the results, and produces a long, fully-cited report instead of a quick answer. OpenAI describes its version as an agent that will "find, analyze, and synthesize hundreds of online sources to create a comprehensive report at the level of a research analyst"¹. Several major AI products shipped broadly similar agentic research features in a tight cluster: Google's Gemini Deep Research first (December 2024²), then OpenAI's ChatGPT deep research (February 2025¹), Perplexity Deep Research (February 2025³), and xAI's Grok DeepSearch with Grok 3 (mid-February 2025), all within about ten weeks; Microsoft's Researcher agent (March 2025⁴) and Anthropic's Claude Research (April 2025⁵) followed that spring. Several use the exact same name, Deep Research.

It resembles agentic retrieval and query fan-out taken to larger scale, though the exact orchestration differs by product: where a quick AI answer fans one query into a handful of sub-queries, a deep research run issues dozens to hundreds of searches over several minutes. For its own products Google states that Deep Search, its AI Mode counterpart, "uses the same query fan-out technique but taken to the next level," able to issue "hundreds of searches"⁶.

Status in 2026

Established as a premium tier across products. The category formed in a first wave between December 2024 and February 2025, and by spring 2025 every engine this glossary tracks had a version: ChatGPT deep research, Perplexity Deep Research, Gemini Deep Research, Microsoft's Researcher (in Microsoft 365 Copilot, March 2025⁴), and Anthropic's Claude Research (April 2025⁵). They are usually gated to a paid plan and run as a multi-minute background task. Note two distinct Google products that are easy to conflate: Gemini Deep Research lives in the Gemini app, while Deep Search is the deeper-reasoning mode inside Google Search's AI Mode. Both are multi-step research surfaces, but they are separate products and should be measured as separate citation surfaces.

How to apply

• Optimize for the sub-questions, harder. A deep research run decomposes a topic into many sub-queries and pulls a passage from whichever source cleanly answers each one. Self-contained, well-sourced coverage of specific sub-intents is more likely to be useful in a multi-source synthesis than head-term ranking.
• Treat it as a separate citation surface. A page cited inside a deep research report is a different observable event from a quick-answer citation; the source pool is deeper and the selection differs, so probe it separately rather than assuming quick-answer results carry over. It is also the most expensive surface to probe (each run takes minutes and consumes paid-tier quota), which is why public citation studies are dominated by quick-answer data and the deep-research citation set is largely unmeasured.
• Expect more slots, deeper competition. Reports cite many sources, so there are more citation slots than a quick answer, but each report draws from a wider pool, so being one of the few clear sources on a specific sub-question is what makes you a likely candidate for a slot.

How it relates to other concepts

• An escalation of agentic retrieval (a multi-step agent loop) that applies query fan-out at much larger scale.
• Distinct from quick-answer AI search, which runs a single-shot fan-out over far fewer sources.
• Its Google form, Deep Search, lives inside AI Mode; it is the deeper sibling of the standard AI Mode answer.