Deep research tools: when a report beats a chat
Deep research tools spend minutes searching, reading and writing a long, cited report. Here’s how they work, how to brief one, and how to check what comes back.
Video transcript
Some AI tools can take a question away, research it for minutes rather than seconds, and come back with a long, cited report. Here’s how to use them well.
It plans the research, runs many searches, reads the pages it finds, and writes it all up with citations. Some can search your email and files too.
All that work takes computing power. Anthropic says multi-agent research systems use about fifteen times as many tokens as a chat. So they’re slower, and many plans limit how often you can run them.
So brief it properly. Give the question and why you’re asking, the scope, time period and region, which sources to prefer or avoid, and the report you want back.
Then check it. Open the key sources. Make sure quotes, numbers and dates match, and that weak sources aren’t doing all the work. Ask what’s missing.
Use it for broad questions that need many sources. For a quick fact or today’s news, an ordinary chat or search is faster and cheaper.
And mind what you upload or connect. It goes to the provider, and a report built from your inbox can quote things you’d rather not share.
Treat the report as a strong first draft, not the final word. The full guide below has a brief you can copy.
In 30 seconds
- A Jargon busterDeep research: An AI feature that plans a research task, runs many web searches, reads the results and writes a long report with citations, taking minutes rather than seconds. tool plans, runs many searches, reads pages and writes a cited report.
- Brief it properly: question, scope, period, region, sources to prefer or avoid, and the output.
- Open the key citations: check quotes, numbers and dates match, and weak sources aren’t doing the work.
Ask an ordinary chat a big question and you get a quick answer from what the model already knows, plus perhaps a search or two. A Jargon busterDeep research: An AI feature that plans a research task, runs many web searches, reads the results and writes a long report with citations, taking minutes rather than seconds. tool takes the question away, runs many searches, reads what it finds and, minutes later, returns a long report full of citations. It’s useful for a broad first map of a topic, and risky if you mistake it for finished work.
What happens while you wait
- PlanSplits your question into smaller ones
- SearchRuns many searches, often in parallel
- ReadOpens pages and files, and follows leads
- WritePulls it into a report with citations
You’ll find it as deep research in ChatGPT, Deep Research in Gemini, Research in Claude and Researcher in Microsoft 365 Copilot, and Perplexity has its own version. Names, limits and features change often, so check what your plan includes.
Under the bonnet, it’s an Jargon busterAI agent: An AI that carries out a task in several steps, such as searching, clicking and filling in forms, rather than just replying., or a team of them, working through the task. That takes far more computing than a chat, which is why these tools are slower, and why many plans limit how many reports you can run.
as many Jargon busterToken: A chunk of text, often part of a word, that a language model reads and writes one at a time. as a chat, roughly: Anthropic’s figure for multi-agent systems like the one behind Claude’s Research.
Source: Anthropic, How we built our multi-agent research system, June 2025Brief it like a researcher
A vague question wastes minutes, and part of your allowance, on the wrong report. Brief it as you would a human researcher.
What goes in the brief
- The question, and whyWhat you need to know and the decision it feeds, so it can judge what’s relevant.
- ScopeWhat’s in and what’s out, such as “home heat pumps, not commercial ones”.
- Time period and regionSay “since 2023” or “the UK”, for example, so you don’t get old figures or another country’s rules.
- SourcesWhich to prefer, such as official statistics and peer-reviewed studies, and which to avoid, such as sellers’ marketing.
- The outputLength, structure, a summary up top, figures in a table, and a list of what it couldn’t find.
Research question: [what you need to know, and the decision it will inform]. Scope: [what to include], but not [what to exclude]. Period: [since 2023]. Region: [UK]. Prefer: [official statistics, regulators, peer-reviewed studies]. Avoid: [sellers’ marketing, forums]. Output: a short summary first, then sections with citations. Put key figures in a table with their source and date. Flag where sources disagree, and list what you couldn’t find.
Check the report before you rely on it
The polish is the trap: headings, tables and dozens of footnotes make a report look checked when it isn’t. Like any AI, it can still make up a fact or misread a source. And when Anthropic built Claude’s Research, its early agents kept choosing pages written to rank well in search over better sources, such as academic papers.
- Open the key citationsCheck each page exists, is what the report says it is, and actually supports the claim.
- Match quotes and numbersFind each quote and figure on the cited page. Watch for wrong years, units, places and rounding.
- Check the datesOld statistics and superseded rules can sit beside new ones. Make sure each source is current.
- Weak sources, repeatedOne blog or seller’s page cited again and again is a red flag, however many footnotes there are.
- Missing viewpointsWho isn’t quoted? Ask for the strongest evidence against the report’s conclusion.
- Misattributed claimsA real source can be cited for something it never says. Check each claim sits with the right link.
Here is a research report. Don’t add new facts. For each key claim, list the citation it rests on and whether that source is official, independent research, news, a seller or opinion. Flag claims with no citation, claims resting on a single source, sources older than [date], and anywhere the report states something more strongly than its source does. [paste the report]
When something simpler is better
| Use deep research | Use something simpler |
|---|---|
| A broad question that needs many sources | A quick fact: an ordinary chat or search is faster |
| An overview of a market, policy or body of research | Today’s news, prices or opening hours: a normal search |
| Comparing many options on set criteria | Questions about your own documents: upload them to a chat |
| A first map of an unfamiliar topic | Your health, money or legal position: a qualified professional |
Running these for a team? Treat every report as a draft: say which tools are approved in your AI policy, and have a named person check the sources before a report is shared or quoted.
Check yourself
3 quick questions nothing is savedTools in this guide
- CChatGPTOpenAI’s general-purpose assistant for writing, questions, analysis and images.
- GGeminiGoogle’s assistant, built into Gmail, Docs and Android.
- CClaudeAnthropic’s assistant, strong at long documents, careful writing and code.
- PPerplexityAn answer engine that searches the web and cites its sources.
- MMicrosoft CopilotMicrosoft’s assistant across Windows, Edge and Microsoft 365 apps.
Sources (3)
- How we built our multi-agent research systemAnthropic, June 2025
- Use research on ClaudeAnthropic (Claude Help Center)
- What is Researcher Agent in Microsoft 365 Copilot?Microsoft (Microsoft 365 documentation on GitHub)
Spotted a mistake? Tell us and an editor will check it.
