It is a way of working where Claude investigates a question across many sources and writes up what it found.

Research is a feature in the Claude apps, described by Anthropic as an in-depth way for Claude to search and reason, delivering thorough answers with citations you can check. Instead of replying straight away from what it already knows, Claude breaks your question into parts, runs a series of searches that build on each other, reads what it finds, and then pulls everything together into an organized report.

The result usually takes minutes rather than seconds. In return, you get broader coverage than a normal reply, and links back to the pages the claims came from. Depending on your setup, Research can look at the open web and, if you connect them, some of your own work sources, such as email, calendar, and documents.

Three levels of effort suit three kinds of questions.

A normal reply comes from what Claude learned in training and the conversation so far. It is fast and is the right choice for explanations, writing help, and stable facts. Web search lets Claude look something up online during a reply, which is ideal for a single current fact. Research goes further. It runs many searches in sequence, explores different angles, and compiles a longer, cited report.

A lead researcher plans, helpers search in parallel, and a final step ties claims to sources.

Anthropic has published a description of how its Research system is built. In plain terms, one lead agent reads your question and makes a plan. It can then start several helper agents that search different parts of the question at the same time. Each helper gathers information and reports back. The lead agent decides whether it has enough, or whether to search more. When it is satisfied, the findings are written into a report, and a separate step works out which source supports which claim, so that citations can be attached.

Because many searches run in parallel and the process can loop until enough is found, a research task takes longer than a normal reply and uses more computing effort. That is why it is useful for open-ended questions where the steps cannot be known in advance.

flowchart TD
    Q["Your research question"] --> L["Lead agent makes a plan"]
    L --> A["Helper 1
searches one angle"]
    L --> B["Helper 2
searches another angle"]
    L --> C["Helper 3
searches a third angle"]
    A --> S["Lead agent reviews findings"]
    B --> S
    C --> S
    S --> D{"Enough information?"}
    D -- "No" --> L
    D -- "Yes" --> R["Write the report"]
    R --> X["Match claims to sources
and add citations"]
        

A few conditions must be true before the feature appears.

Anthropic’s help center describes Research as available on paid Claude plans in the web, desktop, and mobile apps. It has changed over time, starting as an early beta for some plans and regions and expanding afterward, so check the current help center for your plan and country. The help center also states that web search must be turned on for Research to work. You then enable Research from the controls near the message box, or by asking Claude directly to use the research tool.

For Team and Enterprise accounts, administrators may need to enable web search or related connectors for the organization before individuals can use them. If you cannot find the option, the cause is often a plan, region, or admin setting. Menu names and positions change, so rely on the help center for the current steps.

The quality of the report follows the quality of the question.

A research question works best when it is specific, bounded, and clear about what you will do with the answer. “Tell me about electric cars” is too broad and will produce a general overview. “Compare the three most popular compact electric cars for a family of four in a city, focusing on running costs, charging convenience, and safety ratings” gives Claude a target. You can add details that narrow the search: budget, location, time period, audience, and the format you want.

If you are not sure how to frame it, ask for help before starting.

A long research run is a poor time to discover that your question was unclear. A smart habit is to talk it over first in a normal chat. Describe what you are trying to achieve and ask Claude to suggest a sharper research question, point out what you may have missed, or ask you a few clarifying questions. Then use the improved version to start the research. Anthropic’s own guidance for Research includes the idea of asking Claude to help you write a better research prompt before you switch the feature on.

Research takes time, so plan for it.

Once started, a research task can run for several minutes, and complex ones may run longer. Anthropic has described typical tasks finishing in minutes, with some very complex investigations taking considerably longer. You can usually see progress as Claude works, and you can do other things meanwhile. The exact behavior depends on your app and version.

It also helps to know that research uses more effort than a normal conversation. Anthropic has noted that multi-agent research uses many more resources than a simple chat. The help center has a section on how Research counts against usage limits, so if you are on a plan with limits, check how Research is treated and avoid starting many large tasks at once without need.

A good report is organized, sourced, and honest about gaps.

When the research finishes, you will usually get a structured report with sections, key findings, and citations that link to the sources. Read it in layers. First read the summary to see the overall picture. Then look at how the evidence is organized: are the main claims supported by several sources or only one? Finally look for statements about uncertainty, disagreement between sources, or limits of what was found.

A trustworthy report does not hide gaps. If sources disagree, it should say so. If a question could not be answered, it should say that too. If you see confident statements with no citation, treat them with caution, since they may come from general knowledge instead of the research.

flowchart LR
    A["Read the summary"] --> B["Check how claims
are supported"]
    B --> C["Open key sources"]
    C --> D["Note disagreements
and gaps"]
    D --> E["Decide what to trust
and what to verify further"]
        

Citations make checking possible, but you still have to check.

Inline citations let you click through from a claim to the page it came from. That is a major advantage, because you can verify instead of trusting. But a citation is only a pointer. It does not guarantee that the page really supports the claim, or that the page is reliable. Independent reviews of research tools note that citations still need independent verification.

You do not need to check every link. Focus on the claims that matter most to your decision, such as key numbers, quotes, dates, and anything surprising. Open the source, find the passage, and confirm that it says what the report claims. Also look at who published the page and when.

Not every page on the internet deserves equal trust.

Research mode draws from whatever it can find. The open web includes excellent sources and poor ones, and the report is only as good as the pages behind it. You can improve results by asking for source preferences, and you can improve your reading by knowing which sources deserve more trust.

Connecting your tools can make research more relevant, and also raises privacy questions.

Anthropic describes Research as able to search both the web and your connected work context, such as Gmail, Google Calendar, and Google Docs when the Google Workspace integration is connected. This is helpful for tasks such as preparing for a meeting, pulling together action items from email threads, or combining internal documents with outside information. It saves you from uploading files and repeating context.

Connecting these sources also means a research assistant may read private communications and documents. Only connect what you need, understand what each connection can access, and follow your organization’s rules. On work accounts, administrators may control which connections are allowed. Information from your own sources is private to you, so be careful about what you copy into reports that others will read.

The first report is a starting point for a conversation.

After you read a report, you will usually have more questions. You can ask Claude to go deeper on one section, compare two options more closely, check a particular claim, or reformat the findings for a specific audience. Treat the report as a foundation, and use normal chat for the follow-up. Only start another full research run when you need broad new information, since follow-ups in normal chat are faster and lighter.

The right tool depends on the question.

flowchart TD
    A["Your question"] --> B{"One quick current fact?"}
    B -- "Yes" --> C["Use web search"]
    B -- "No" --> D{"Needs many sources,
comparison, or a report?"}
    D -- "Yes" --> E["Use Research"]
    D -- "No" --> F["Use a normal reply"]
        

Research mode is powerful, but it has real boundaries.

Using research responsibly protects you and other people.

Be thoughtful about what you ask and how you use the results. Avoid sending confidential or regulated data into a research request unless your organization has approved it. Do not use research to compile private information about individuals, such as where a private person lives or works. When you quote or build on what the report found, credit the original sources, and do not present Claude’s summary as your own original investigation.

If you use research in academic or professional work, check the rules about disclosing AI assistance, and verify everything yourself. A report that Claude produced is a draft to examine, not a finished, citable authority.

Familiar jobs show how the pieces fit together.

Research saves effort, but it asks for care in return.

These habits cause most disappointing research results.

Follow these steps for dependable results.

A few popular beliefs about research mode are only partly true.

Quick answers to the questions beginners ask most.

It is a feature where Claude plans a search, looks at many sources, and writes a cited report. It takes longer than a normal reply but covers more ground.

Make sure web search is on, then enable Research from the controls near the message box or ask Claude to use the research tool. Menu names change, so check the current help center.

Anthropic’s help center describes it as available on paid plans, and availability has expanded over time. Confirm for your plan and region in the current help center.

Often minutes, with more complex questions taking longer. You can usually do other things while it runs.

If you connect supported tools such as Google Workspace, Research can include them alongside the web. Connect only what you need and follow your organization’s rules.

State the goal, the scope, constraints, the criteria for comparison, the audience, and the format, and ask Claude to help refine it first.

They make checking possible, but you should still open the sources for key claims, confirm they say what is claimed, and judge the quality of each source.

For a single quick fact, a stable concept, or writing from your own material, a normal reply or web search is faster and lighter.

Research uses more effort than a normal chat, and the help center explains how it counts toward usage limits, so check it for your plan.

No. It cannot reach all databases or follow formal review standards, so treat it as a starting point and an aid to exploration.

Verify the key claims, ask for counterarguments and unknowns, reshape the output if needed, and confirm any high-stakes decision with a qualified professional.

Research mode lets Claude plan a search, explore many sources in parallel, and return an organized report with citations. It is best for broad questions that need comparison, current information, or background from many places, and it is not needed for quick facts or writing tasks. You get the most from it by asking a sharp question with a clear goal, scope, criteria, and format, by waiting patiently, and by reading the report in layers. Citations make checking easy, but you still need to open the sources that matter and judge their quality. Be careful with connected work sources, keep sensitive data protected, confirm high-stakes decisions with experts, and recheck the help center because plans and features change.