Web search and Research

You are here · All paths · Free plan has limited search; Research needs Pro · 40–50 min · Assumes Prompting fundamentals.

What you’ll learn: when to use plain search vs. Research mode, and how to get output you can actually trust.

If you only read one thing

Claude’s built-in knowledge stops at a fixed date. For anything current, it has to look things up, and there are two ways it does that.

Web search is a handful of quick lookups with citations — right for a fact, a price, a recent announcement.

Research is a different animal: Claude plans an investigation, runs many searches, follows leads, and writes you a sourced report. It takes minutes rather than seconds. Right for “what’s the landscape here”, not “what’s the capital of Peru”.

The discipline that matters in both: click the citations. A confident summary attached to a source that doesn’t say that is the most common way people get burned. Spot-check two or three links on anything you’re going to act on.

Three different things

They compose. Research uses search and thinking.

Web search

Claude decides when to search based on your question, or you can force it. Results come back with citations you can click.

Always search for:

• Anything after the model’s knowledge cutoff (May 2026 for Opus 5, Jan 2026 for Sonnet 5 and Fable 5)
• Prices, versions, release dates
• Who currently holds a role
• Whether something is still true
• Anything where being out of date would be embarrassing

Prompting search well:

Be specific about recency and source quality.

Bad: what's the best vector database

Good: What are the leading vector databases as of mid-2026? Prioritise sources from the last six months. For each, note pricing model and whether it supports hybrid search. Cite everything.

Verify the citations

This matters. Claude cites what it retrieved, but summarisation can drift from the source. For anything consequential, click through on at least the load-bearing claims. The failure mode isn’t fabricated URLs any more — it’s a real source being characterised slightly wrong.

Ask for it explicitly:

For each claim, quote the exact sentence from the source that supports it.

Research

Research runs a longer, multi-step investigation. Claude plans an approach, searches repeatedly, follows leads, reconciles conflicting sources, and returns a structured report with citations. It can also search your connected apps (Gmail, Drive, etc.) if you have connectors enabled.

Use it for:

• Competitive landscapes
• Literature and prior-art surveys
• “What’s the current state of X?” where X is broad
• Due diligence
• Anything where you’d otherwise open thirty tabs

Don’t use it for: single facts. It’s slower and it’s overkill.

Getting a good report

Research responds enormously to scoping. Give it:

1. The question, sharply. “Should we build or buy an auth system?” not “tell me about auth.”
2. Your context. Team size, constraints, what you already ruled out.
3. What “done” looks like. “A recommendation with three options compared on cost, time-to-ship, and compliance burden.”
4. Source preferences. “Prioritise primary sources and vendor docs over listicles.”
5. The output shape. “A table comparing options, then a one-page recommendation.”

Example:

Research the current landscape for European payment processors suitable for a B2B SaaS doing ~€2M ARR, selling into Germany, France and the Netherlands.

We need: SEPA direct debit, invoice payments, and card. We already ruled out Stripe on pricing.

Deliver: a comparison table (pricing, supported methods, SEPA support, integration effort, notable limitations) covering at least five providers, then a one-page recommendation with reasoning. Prioritise vendor documentation and recent independent reviews over marketing content. Cite everything.

The structured research prompt

For genuinely hard research — the kind where you want Claude to be rigorous rather than fast — this pattern (from Anthropic’s own guidance) works well:

Search for this information in a structured way. As you gather data, develop several
competing hypotheses. Track your confidence levels in your progress notes to improve
calibration. Regularly self-critique your approach and plan. Update a hypothesis tree
or research notes file to persist information and provide transparency. Break down this
complex research task systematically.

It changes the character of the output from “here’s what I found” to “here’s what I found, here’s what contradicts it, and here’s how confident I am.”

Trust calibration

A practical hierarchy for how much to trust research output:

Try it

Exercise 1 — Cutoff demonstration. Ask about something that happened last month with search off, then on. Note what the model does when it doesn’t know.

Exercise 2 — Vague vs. scoped research. Run Research twice on the same topic: once with a one-line question, once with the full scoping structure above. Compare the reports.

Exercise 3 — Citation audit. Take any Research report and check five citations at random. Rate each: accurate / slightly off / wrong. This gives you a personal calibration you’ll rely on for years.

Exercise 4 — Competing hypotheses. Run a research task with the structured prompt above. Note whether it surfaces contradictions it would otherwise have smoothed over.

Exercise 5 — Real decision. Use Research on an actual decision you’re facing. Judge it by whether it changed your mind about anything, not by whether it sounded good.

Checkpoint

• You know when search is enough and when Research earns its extra minutes
• You have a personal sense of how often citations need checking
• Your research prompts specify output shape and source preferences

Going deeper

• When should I use web search, extended thinking, and research?
• Enable and use web search
• Use research on Claude
• Reduce hallucinations