Phase 3 · Claude · Level 2 · Practitioner
Web search and Research mode: cited, multi-source answers
By the end, you'll be able to…
- Tell the difference between plain web search and Research mode, and pick the right one
- Run a Research task that returns a cited, multi-source answer on a real work question
- Verify citations properly instead of trusting a report because it has footnotes
Why it matters
Claude learns from information up to a knowledge cutoff, so on anything recent it can be out of date. Web search fixes that for quick questions; Research mode goes further, working through many sources to build a cited report. Both are powerful and both can mislead if you trust the citations without opening them. This lesson shows when each is worth it, and how to check the sources so a tidy report doesn't fool you.
Why this exists: the knowledge cutoff
A model like Claude is trained on information up to a point in time, its knowledge cutoff. Past that point it doesn't inherently know what happened: a policy that changed last month, a price updated last week, a competitor's announcement yesterday. Ask about any of those from memory and you risk a confident, outdated answer.
Connecting Claude to the live web solves this, and it comes in two strengths: web search for quick, current questions, and Research mode for deep, multi-source work. Knowing which to reach for saves both time and trust.
Web search: quick and current
With web search enabled, Claude can look things up as it answers, fetching current pages, then replying with links to where it found things. It's the right tool for a single fact or a fast-moving topic: "what's the current UK VAT registration threshold?", "has this supplier announced anything this quarter?". Claude searches, reads, and answers with sources you can click.
Search the web: what is the current UK VAT registration threshold, and from what date does that figure apply? Give me the exact figure, the effective date, and a link to the official government source you used.
Why this works: Asking for the figure, the date it applies from, and the source link turns a lookup into something checkable: you can open the link and confirm rather than trusting the number on its own.
The reply is only as good as its source, which is why asking for the link, ideally an official one, is not optional. Web search reduces the out-of-date risk; it doesn't remove the need to check.
Research mode: deep and multi-source
Research mode is a different gear. Instead of a quick lookup, Claude works agentically: it plans an approach to your question, breaks it into parts, runs many searches (often in parallel), reads full pages rather than snippets, and pulls the findings together into a structured, cited report. It can take several minutes, and where you've connected sources like email or documents, it can weave those in alongside the web.
You start it from the plus (+) button in the chat and choosing Research; web search must be on for it to work. Use it when a question needs cross-referencing several sources: market or competitor analysis, a literature or policy review, due diligence, comparing options across vendors. For "what's the VAT threshold", it's overkill. For "how are mid-size UK firms handling hybrid-working policy in 2026, and what do the main viewpoints disagree on", it's exactly right.
Research how mid-size UK companies are approaching hybrid-working policy in 2026: typical office-day expectations, common core-hours arrangements, and the main points of disagreement between employers and staff. Give me a structured summary with citations, and explicitly flag where your sources conflict or where the evidence is thin.
Why this works: A good research prompt names the angle, the scope and the output shape, so Claude's plan targets the right sources and returns something structured. Asking it to note where sources disagree is what makes a research report more valuable than a single search.
What comes back reads like a briefing: sections, claims, and citations you can follow. The temptation is to take it as finished. Don't. The citations are the beginning of your job, not the end of Claude's.
Verifying citations: the part that matters
A cited report looks authoritative, and that's precisely the risk. Two failure modes are worth naming. First, a citation might not actually support the claim attached to it. The source is real, but it says something subtly different, or the claim overstates it. Second, on the open web, a confidently-cited "fact" can trace back to a weak or biased source dressed up as authority.
So verify like this:
- Open the load-bearing links. For any claim you'll act on or repeat, click through and check the source really says what Claude claims. Don't verify only that a link exists; check that it supports the point.
- Judge the source, not just its presence. An official body, a regulator, a primary document beats a random blog. A footnote to a content-farm article is not the reassurance it looks like.
- Watch for a claim with no citation, or one citation stretched across several claims. Those are the seams where errors hide.
- Cross-check anything surprising. If a finding is more convenient or more dramatic than you expected, that's the one to confirm from a second independent source.
For the research above, split your findings into two lists: claims backed by strong, official or primary sources, and claims resting on weaker or single sources. For each strong claim, give the direct link. For each weak one, say what a better source would be. I'll verify the weak ones myself.
Why this works: Asking Claude to separate well-supported points from thin ones, and to rank source quality, does some of the triage for you, but you still open the links. It focuses your checking on exactly the claims most likely to be shaky.
Try it now
Common mistakes
- Using Research mode for everything. It's slower and heavier than it needs to be for a quick fact. Match the tool to the question: web search for lookups, Research for real cross-referencing.
- Treating citations as proof by their mere presence. A footnote means a source was consulted, not that it supports the claim. Open the ones that matter.
- Ignoring source quality. A cited blog isn't a cited regulator. Weigh where the claim comes from, not just that it came from somewhere.
- Forgetting even fresh answers can be wrong. Web search reduces the knowledge-cutoff problem; it doesn't remove misreading, or a bad source read correctly.
- Over-trusting a polished research report. The more structured and cited an answer looks, the more it feels done, and a confident report built on a couple of shaky sources is more dangerous than an obvious guess, precisely because it disarms your scepticism. Verify the load-bearing claims yourself before any of it informs a decision.
Keeping current
Web search and Research mode are evolving fast: more connected sources, better planning, richer citations. The durable skills (match the depth to the question; open the links; judge the source) will outlast any interface change. Check the official Research help article and the Claude release notes for what's current. Accurate as of 13 July 2026.