Claude Research: When a Multi-Step Search Is Worth It
6 min read
A product manager needs to know how three competitors price their entry plans, what customers complain about most, and whether any of them changed their terms this year. She could ask Claude one question and get a decent paragraph. Or she could ask for Research, wait while Claude runs a chain of searches that each depend on what the last one found, and get something closer to a briefing. Which of those she should choose depends on the question, not on how important it feels.
This article explains what Research does, what it needs, and how to write the request so the extra effort pays off. It builds on When Claude Is Searching the Web and When It Isn't, which covers the single-search case.
What Research actually is
Anthropic describes Research as Claude investigating a question in depth by running multiple searches that build on each other. It decides what to look into next based on what it has already found, explores different angles, and works through open questions systematically. A normal web search answers a question. Research tries to build an understanding of a topic.
A few facts from Anthropic's help page shape how you use it:
- It is available on paid plans only: Pro, Max, Team or Enterprise.
- It works on the web, in Claude Desktop and in Claude Mobile.
- Web search has to be on first.
- You turn it on from the "+" button at the bottom left of the chat, by selecting "Research." A blue indicator appears at the bottom of the window, and clicking it turns Research off.
- It can use your connectors, including Google Workspace (Gmail, Google Calendar and Google Docs), when those are turned on for the chat, and it uses connector tools automatically without asking for approval each time.
That last point deserves a second look. Research can reach into your email and documents without a confirmation step for every action. Turn on only the connectors a given question needs.
Two quiet limits
Research started in Claude Mobile searches only the web and skips connectors, with no error or notice. If your question depends on your own documents, start it on the web or in Claude Desktop. Research also cannot use tools from local MCP servers configured in Claude Desktop.
Which model it runs on
Anthropic says Research runs on Claude Opus 5.5 and falls back to Claude Opus 4.8 when needed, whichever model you picked for the chat. On Team and Enterprise plans, if an admin has turned off both of those models for your role, Research will not work for you. If it is missing and you expected it, that is the first thing to ask your admin.
It follows the same limits, and uses them faster
Research follows the same usage limits as standard conversations, but it can consume them faster because it retrieves many sources and returns a comprehensive response. Anthropic's page does not state how long a run takes or whether the output follows a fixed citation format, so plan around that uncertainty: start a Research run when you can wait, and check the sources yourself before you rely on the result.
Pick the right tool for the question
A normal answer or one search is enough
- One fact that can be checked in a single source
- A document you already have: summarize it, do not research it
- Rewriting, structuring or brainstorming
- A question with a short, stable answer
Research earns its cost
- Several sub-questions that depend on each other
- A comparison across multiple sources or companies
- A topic you cannot yet describe well enough to search for directly
- A decision where a wide, sourced overview beats a quick answer
Write the brief before you press go
A vague request produces a long, vague report. The best habit is to write a short brief first. Five lines are enough.
- 1
State the decision, not just the topic
"I am deciding whether to price our entry plan below competitors" is a brief. "Competitor pricing" is a topic. The decision tells Claude what is relevant and what to leave out.
- 2
List the sub-questions
Name the two to four things you need answered. Research handles sub-questions well, because that is how it works through a topic.
- 3
Say what counts as a good source
"Prefer official pricing pages and company announcements. Treat forum posts as opinion, not fact." Without this, Claude weighs every source equally.
- 4
Set the boundaries
Time window, region, company size, language. "Pricing as of this year, for customers in Europe" removes a lot of noise.
- 5
Ask for the shape of the output
"A one-page summary, then a table of findings with a source for each row, then a list of what you could not confirm." The last item is the most valuable, because it shows you where to look yourself.
A complete request looks like this:
I am deciding whether to price our entry plan below our three main competitors. Use Research. Find: (1) the current entry-plan price and what it includes for each of A, B and C, (2) any change to their terms or limits this year, (3) the most repeated complaints in their public reviews. Prefer official pricing pages and announcements, and treat forum posts as opinion. Customers are in Europe. Give me a one-page summary, then a table with one source per row, then a list of anything you could not confirm.
”If Research doesn't start
Sometimes Claude answers normally instead. Anthropic's advice is to prompt it directly, for example: "Please use the research tool to answer this." If it still does not run, check that web search is on and that your plan includes Research.
Read the report like an editor
A research report reads smoothly, which is exactly why it needs checking. Three habits catch most problems.
- Open the sources behind the numbers. Pick the three claims that would change your decision and check each one against the page.
- Look for what is missing. If you asked for three competitors and the report is thin on the third, say so and ask Claude to look again at that one.
- Separate findings from inference. Ask: "Which statements here come directly from a source, and which are your interpretation?" The split changes how much weight each sentence deserves.
Tip
Use Research to get a map, not a verdict. The map tells you where the answers are. You still decide which ones to trust, because you are the one who carries the consequences.
If you also use other assistants, ChatGPT Deep Research: When It's Worth the Wait and Writing a Deep Research Brief Gemini Can Actually Use make the same case for their own versions, and the brief structure above works with all three.
Official sources
Checked on October 24, 2026. Features, plans and names change often, so the vendor's own pages are the final word.