Claude Research: how Claude's deep research works, and how to check its report

A strategy analyst has until Thursday to brief her director on the market for commercial heat pumps in the UK: who the main suppliers are, what the subsidy schemes pay, where the growth is. In an ordinary chat with web search on, Claude runs a search or two and answers. In Research mode, it plans, searches, reads what it finds, decides what to look into next, and searches again, then writes a report with links to its sources.

Anthropic's description: Research "runs multiple searches that build on each other and decides what to look into next based on what it finds" (Anthropic Help Centre, Using Research). Three details matter more than the marketing suggests:

• It searches your apps too. Any connectors you've enabled, such as Google Workspace, are searched alongside the web. On mobile it only searches the web, and gives no sign that your connectors were skipped.
• It picks its own model. Research runs on Claude Opus 5.5 (falling back to Opus 4.8) regardless of the model you selected.
• It takes a while. Most reports arrive in 5 to 15 minutes; complex ones can run for up to 45 (Anthropic, May 2025).

Starting a session is simple:

1. Make sure web search is on. Research depends on it.
2. Select "+" (bottom left), then Research. A blue indicator shows it's active.
3. Turn off connectors that can write, such as anything that sends messages or edits records. During Research, tools can run without asking you first, and Anthropic advises disabling write-capable ones (Anthropic Help Centre, Custom connectors).
4. Write the brief, and send it. Then do something else for a quarter of an hour.

The brief decides the report. Anthropic's prompting guide for research tasks recommends clear success criteria, asking Claude to verify information across multiple sources, developing several competing hypotheses, and tracking its confidence (Anthropic, Prompting best practices). Put together, a brief looks like this (illustrative):

I'm briefing a director on the UK commercial heat pump market, 2023 to now.
I need: the five largest suppliers by UK installations, current government
grant schemes and what they pay, and forecasts of market growth.

Prefer primary sources: government publications, company filings, trade body
data. For every figure, give the source, its publication date, and whether
you found the same figure in a second independent source.
Where sources disagree, show both figures; don't average them.
End with a list of what you searched for and couldn't find.

The last line is the most useful one in the brief. A model that can't find something tends to fill the gap; asking for a list of gaps turns them into information.

One cost to know: research sessions run under the same limits as normal chats, but Anthropic says they "can use up your limits faster".

Research also produces a lot of claims at once, and its citations can make a report look more checked than it is. Count connects to Claude and splits the report into claims, each tagged as a citation, an assumption, model knowledge or a hallucination.

Research is at its best on breadth. Mapping who the players are, what has been published, which questions have consensus and which don't: a first pass that would take a person a morning of tabs. Combining public sources with your own documents through connectors is something an ordinary web search can't do at all.

It's weakest where a single precise fact matters. A market-size figure, a regulation's effective date, a competitor's price. Those tend to come from secondary articles that quote other articles, and a long report gives each one less scrutiny than a short answer would. Anthropic describes Research's citations as "easy-to-check", which is true. Easy to check is not the same as checked.

We'd use Research to decide what's worth reading, then read it. For a head-to-head with OpenAI's equivalent, see Claude vs ChatGPT for research.

We found no independent evaluation of Claude Research's citation accuracy specifically. The studies of the category are consistent, though, and not flattering.

• Citation accuracy of 40–80%. A Salesforce study of generative search and deep-research systems found large fractions of statements weren't supported by the very sources cited next to them (Venkit, Laban et al., DeepTRACE, September 2025).
• Working links, wrong claims. Across 14 models, the best produced valid links more than 94% of the time, yet only 39–77% of claims were factually accurate. Accuracy fell by about 42% as tool calls grew from 2 to 150 (Onweller et al., Cited but Not Verified, May 2026). A Research session is exactly that kind of long tool run.
• Confident when wrong. When Columbia's Tow Center tested eight AI search tools on 200 news excerpts, ChatGPT Search got 134 wrong and declined to answer none of them (Tow Center, March 2025). That's search, not deep research, and a different vendor. The pattern of fluent, unhedged error is the one to watch for.

A two-minute check on a Research report

1. Find the three claims the conclusion depends on. Not the first three: the ones that would change the recommendation if wrong.
2. Open their citations and find the sentence that supports each claim. If you can't find it in a minute, mark the claim as unsupported.
3. Check what kind of source it is. Primary (the filing, the statistics release) or an article about it? Numbers drift with each retelling.
4. Check the dates. A 2022 forecast quoted as current is a common slip in fast-moving markets.
5. Ask Claude: "Which claims in this report rest on a single source? What did you search for and not find?"

Doing that for every claim in a 3,000-word report is a lot of clicking. Count does it from inside the conversation: connect it to Claude, ask it to audit the report, and each claim comes back labelled as a citation, an assumption, model knowledge or a hallucination, with the reasoning chain showing which search it came from.

Research launched in April 2025 as an early beta for Max, Team and Enterprise users in the US, Japan and Brazil, with Google Workspace search (Anthropic, April 2025). The longer "advanced" mode, up to 45 minutes, arrived on 1 May 2025, and Pro users got access on 3 June 2025 (Anthropic, May 2025).

A Research report is a well-organised reading list with a draft conclusion on top. The structure is valuable. The conclusion is the part to hold loosely, because synthesis is where errors hide: every individual citation can check out while the argument that strings them together doesn't.

The risk grows with polish. We've written about hallucinations built from true facts and a wrong conclusion, and about how a confident explanation can stop readers checking at all. A 20-page cited report is about as confident and fluent as AI output gets. Check the claims the decision rests on, and let the rest be what it is: a fast, useful first draft.

Does Claude have deep research?

Is Claude Research free?

How long does Claude Research take?

Is Claude Research accurate?

Which model does Claude Research use?