Deep research

Deep research is an AI agent capability in which the agent autonomously plans a multi-step research task, runs many searches, reads the resulting sources, keeps notes, and synthesizes a long-form, cited report — the kind of work that used to require an analyst spending hours or days pulling material together. OpenAI, Anthropic, Google, and Perplexity each ship a version of this under a similar name, and while the implementations differ in detail, they share the same core pattern of planning, searching, reading, and synthesizing in a loop rather than a single pass.

What separates deep research from an ordinary AI-assisted search is time and depth: a normal query returns an answer in seconds from a handful of sources, while deep research is explicitly built to spend minutes running dozens of searches and reading dozens of sources before producing anything. That tradeoff only makes sense for questions where thoroughness matters more than speed, which is precisely the kind of task deep research has been adopted for across research, sales, and analysis work.

How deep research works

The process starts with a plan: the agent breaks the user's question into a set of sub-questions, decides which types of sources to prioritize, and often notes what falls out of scope.

It then moves into a search phase, running agentic search across many queries in parallel and in sequence, against the open web, internal databases, or both.

As results come back, the agent reads and takes notes, fetching sources, extracting the passages that matter, and maintaining a running set of findings tied back to their citations.

Gaps in those notes trigger further searches, and contradictions between sources prompt a deeper read rather than picking one source arbitrarily — this iteration step is what distinguishes deep research from a single retrieval pass.

The final step is synthesis, where the agent writes a long-form response, often running to thousands of words, with inline citations tracing every claim back to a specific source.

Deep research versus a normal search query

A normal query runs one search, returns a handful of top-ranked results, and produces a brief synthesized summary in a few seconds. Deep research runs dozens of searches, reads dozens of sources, takes anywhere from several minutes to tens of minutes, and produces a structured, multi-section report rather than a short summary.

The difference is not just scale but intent: a normal query is optimized to answer quickly and adequately, while deep research is explicitly optimized for coverage and accuracy at the cost of latency, which only makes sense when the question genuinely warrants that investment.

What makes deep research useful in practice

Coverage is the first advantage — reading 30 sources in eight minutes is faster than almost any human could manage, and the agent has no tendency to skip a source just because it looks tedious.

Citations are the second advantage: every paragraph in the output traces back to a specific URL, which lets a reader verify a claim rather than simply trust it.

Structure matters too, since a well-organized report with clear sections, comparisons, and supporting detail is considerably more usable than a raw list of links a person still has to synthesize themselves. And the whole approach depends on being explicitly tuned for depth over speed — deep research is expected to take real time, which is a deliberate design choice rather than a limitation to work around.

Where deep research runs into trouble

Source quality sets a hard ceiling on output quality: deep research is only as good as what it reads, and poor source selection is the most common failure mode, since grounding a report in weak sources just produces a well-organized version of the same weak information — see AI grounding for the underlying concept.

Recency is a second limitation, since web indexes and provider caches lag behind real time, which means deep research can confidently cite sources that are already stale for breaking news or fast-moving markets.

Bias amplification is a subtler risk — if the agent's initial research plan leans in one direction, its follow-up searches tend to reinforce that lean rather than correct for it, which puts real weight on how the initial plan is framed.

Cost is the final constraint: a single deep research run can cost ten to a hundred times a normal query, given the dozens of underlying model calls, source fetches, and a long final generation step.

Where deep research fits in an AI stack

Enterprise teams use deep research internally for tasks like account research ahead of a sales call, competitive analysis, market briefings, or litigation research — anything where the value of thoroughness clearly outweighs the cost of waiting several minutes for an answer. In a customer-experience context specifically, a deep-research agent can run against internal knowledge bases and the public web to answer complex comparative questions, or to prepare a briefing for a human specialist ahead of a handoff.

Underneath, deep research relies on the same primitives as any agentic system — planning, tool calling, and iterative retrieval — just tuned for thoroughness instead of speed. As support organizations push AI further into complex, judgment-heavy requests, deep research is the pattern that lets a single agent match the depth of research a skilled human specialist would otherwise have to do by hand before responding.