Deep research is an AI feature that turns one question into a small research project: the AI makes a plan, runs many web searches, reads the pages it finds and, after several minutes, writes a long report with links to its sources. Google Gemini (Deep Research), Claude (Research) and Perplexity (Deep Research) all offer a version.
A normal chatbot answer is like asking a librarian a quick question at the desk. Deep research is like asking them to spend the afternoon in the stacks and come back with a written briefing and a list of references. The briefing saves you hours, but you still look up the references that matter before you rely on them.
In a sentence: “I gave deep research my question about heat pumps in old houses, checked its plan, and a few minutes later had a long report with a list of sources to work through.”
How is deep research different from a normal chatbot answer or a search?
The plain point: it trades speed for breadth, and hands you more to check, not less.
| Tool | What you get |
|---|---|
| Search engine | Differs: a list of links in a second or two. You do the reading. |
| AI search engine | Differs: a short written answer from a few pages, with links, in seconds. |
| Deep research | Differs: a long report from many pages, with links, in minutes. The AI does many rounds of searching and reading first. |
| Reasoning model | Differs: a model that thinks longer before answering. Deep research is a feature built on top, adding the searching and the report. |
How does deep research work?
The plain point: plan, search, read, repeat, then write it up.
Google says Gemini’s version “transforms your prompt into a personalized multi-point research plan”, that “You’re in control of the plan”, and that it can “automatically browse up to hundreds of websites” before writing a report.1 Anthropic says Claude’s Research feature “operates agentically, conducting multiple searches that build on each other” and delivers answers “complete with easy-to-check citations”.2 Perplexity says its version “performs dozens of searches, reads hundreds of sources, and reasons through the material”.3
Behind the scenes it can be a team of AIs. Anthropic describes its system as “a lead agent” that delegates to helper agents (subagents) searching “in parallel”, with a final step that matches each claim to a citation.4 That is why it is called an agentic feature: the AI decides its own next search.
What can go wrong with deep research?
The plain point: more sources is not the same as better sources.
The AI picks which pages to trust, and it can pick badly. Anthropic says its early research agents “consistently chose SEO-optimized content farms over authoritative” sources until it added rules about source quality.5 A content farm is a site that churns out pages written to rank in search, not to be right. Google’s own page for the feature says “Check responses”.1
It is also heavy on computing. Anthropic measured that agents “typically use about 4× more tokens than chat interactions”, and its multi-agent setups about 15× more.5 A token is a small chunk of text the model reads or writes. So it suits questions worth the wait, not quick look-ups.
Deep research is a fast first draft of a literature search: a plan, many pages read, a report with links. Read the plan before it starts, and open the sources behind any claim you will act on.
Questions people ask
How long does deep research take?
Minutes rather than seconds. Perplexity says its version finishes “in 2-4 minutes”, and Google says Gemini’s reports are “generated in minutes”.31
Is deep research accurate?
Not guaranteed. It can lean on weak sources, as Anthropic found in its own agents, so treat the report as leads to check; how to check an AI citation takes a minute per link.5
Is deep research the same as a reasoning model?
No. In this site’s reading, a reasoning model is the engine that thinks before answering; deep research is a feature that sends a model out to search and read, again and again, before it writes.Reasoning, 19 Oct 2026 — this site’s distinction, based on the vendor descriptions cited above.
Where to go next
- Checking what it cites: how to check an AI citation.
- Searching well: how to search now.
- Why AI gets things wrong: why AI makes things up.
- The idea underneath: what is an AI agent? and grounding.
- The words: the AI glossary.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What does a deep research feature do before writing its report?
- It makes a plan and runs many rounds of searching and reading
- It asks a human expert to check the answer
- It answers from memory only, without searching
Show the answer
Answer: It makes a plan and runs many rounds of searching and reading. The page says deep research plans, searches, reads and repeats before it writes a long report with links.
2. What problem did Anthropic find in its early research agents?
- They refused to search the web at all
- They chose content farms over authoritative sources
- They wrote reports without any links
Show the answer
Answer: They chose content farms over authoritative sources. The page quotes Anthropic saying its early agents picked SEO-optimised content farms until it added rules about source quality.
3. How does the page suggest you treat a deep research report?
- As a finished answer that needs no checking
- As something to use only for quick look-ups
- As a first draft whose key sources you open and check
Show the answer
Answer: As a first draft whose key sources you open and check. The takeaway calls it a fast first draft: read the plan, and open the sources behind any claim you will act on.
Related terms: Context engineering · Context rot · Context window · Few-shot prompting · Grounding · Knowledge cutoff
- 1Google, Gemini Deep Research, read at source 19 Oct 2026Deep Research is “an agentic feature in Gemini that can automatically browse up to hundreds of websites”; it “transforms your prompt into a personalized multi-point research plan”; “You’re in control of the plan.” ↑a“Check responses”. ↑breports are “generated in minutes”. ↑c
- 2Anthropic, Claude takes research to new places, 15 Apr 2025, read at source 19 Oct 2026“Claude operates agentically, conducting multiple searches that build on each other”, “It explores different angles of your question automatically”, and “delivers thorough answers, complete with easy-to-check citations”. ↑
- 3Perplexity, Introducing Perplexity Deep Research, 14 Feb 2025, read at source 19 Oct 2026“performs dozens of searches, reads hundreds of sources, and reasons through the material”. ↑acompletes “in 2-4 minutes what would take a human expert many hours.” ↑b
- 4Anthropic, How we built our multi-agent research system, Anthropic Engineering, 13 Jun 2025, read at source 19 Oct 2026“a lead agent coordinates the process while delegating to specialized subagents that operate in parallel”, and the system “passes all findings to a CitationAgent” to “identify specific locations for citations.” ↑
- 5Anthropic, How we built our multi-agent research system, 13 Jun 2025, read at source 19 Oct 2026early agents “consistently chose SEO-optimized content farms over authoritative” sources. ↑a“agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats.” ↑bearly agents “consistently chose SEO-optimized content farms over authoritative” sources. ↑c
The comparisons and examples on this page (the librarian, the heat-pump question) are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK