Web search and Research

Open claude.ai in your browser and check the selected account before practicing. This chapter uses the Web interface. Control placement can change; local files are available only after upload or a supported connection. This exercise does not require installing Desktop.

By the end, you will frame a researchable question, obtain sources and check whether they support the answer. Example outputs are templates, not completed research findings.

Web search fits a focused question needing current information. Reading a link fits a particular document. Research makes a broader investigation through searches that build on one another. Thinking helps reason about the problem but is not an external information source.

Sources checked on September 30, 2026 describe Research on paid plans, including Desktop. New interfaces may search without a separate Web search switch; older ones use +. A missing switch does not prove search is unavailable.

“Research learning” is too broad. Connect the question to a decision:

I am planning a fictional internal course for adult employees.
Search the web for evidence about retrieval practice and spaced repetition
compared with rereading alone.
Separate research findings from suggestions for course design.
Prefer research reviews and primary studies over vendor blog posts.
Give each source's date and link, study limits and unsupported conclusions.

This defines the audience, comparison and evidence type. It does not authorize choosing a budget, vendor or organizational policy. Add time, subject-area and accessibility constraints when relevant.

Check that search is available, then ask:

Use web search to find one research review on retrieval practice.
Return its title, authors, year, link and a cautious finding summary.
If the full text is unavailable, say so. Do not claim to have read it.

Watch for search activity and sources. Open the link, check the title and year, and read the passage supporting the claim. A search result is not necessarily a fully read source.

A useful output structure, deliberately without invented findings:

Source: [actual title], [year], [link that opens]
Finding: [claim supported by the source]
Limit: [population, study conditions or access limitation]
Does not establish: [broader unsupported conclusion]

Fill these fields only after inspecting a real source. Do not copy the placeholders into a report as findings.

If Research is available, select it through + or the corresponding control. The official guide describes dependence on web search; an organization may need an administrator to enable the capability.

Use Research to compare retrieval practice, spaced repetition and rereading
for a fictional adult-employee course.
Find at least four relevant sources, including a review and a primary study
if available. For each: source type, year, population, finding, limit and link.
Check whether sources disagree or study different conditions.
End with two options for a small course trial, not a declaration of a winner.
Do not use email, calendar or private documents for this task.

Research can use connected-service context. Say whether the task is web-only or includes internal material. A connected service is not a reason to include unrelated private information.

Without Research, run focused searches, preserve sources and synthesize after verification. Do not call that a Research-mode run if you did not use the mode.

Choose two important claims:

For each report claim, identify the supporting source and what it says.
Mark interpretations and design proposals separately from direct findings.
Do not count a link that fails to open as verified evidence.

Inspect these points yourself:

• The link reaches the intended document, not a homepage or similar result.
• The publication year is correct. Newer is not automatically better.
• The study population fits the question, or the gap is stated.
• The conclusion is no stronger than the finding. Correlation is not causation.
• Reviews, primary studies, marketing and community accounts are labeled separately.

For inaccessible material, state the limit and find an alternative or available abstract. Do not invent quotations from unread papers or bypass paid access.

After checking sources:

Using the verified sources, propose a small fictional course trial.
One group receives short retrieval questions; another rereads a summary.
Define what to measure, when to measure it and what to keep the same.
Label this a trial proposal, not a promise of improvement.
Do not invent results or approval to conduct the trial.

Useful output separates evidence from design choices: question type, comparison conditions and measurement. It does not promise a 30% improvement without suitable evidence. Before real use, check workplace requirements, accessibility and any required consent.

A final table can separate what is known, what is limited, what is proposed and what needs a decision. A long generated report is not automatically the right handoff; ask for a decision summary with sources readers can open.

Easy: one accessible source

Find and open a source. Success: title, year and the claimed finding match the document.

Intermediate: source types

Compare a research source with a training vendor's post. Success: distinguish evidence from an implementation proposal or sales claim.

Challenging: a bounded conclusion

Check whether a report conclusion exceeds its evidence. Success: revise it to state population, conditions and uncertainty while keeping practical value.

No search switch: request search explicitly and check that it happened. Also check organizational policy.

Many links, little evidence: narrow to one question and request claim-level support. Source count alone is not quality.

Research is slow or uses more allowance: it searches and reads more. Match scope to the question.

Long linked document: request the relevant section. Reading everything can consume substantial context and usage.

Wrong local results: specify the city or area; search may use approximate IP location.

Search brings sources. Your job is to check that they are real, relevant and support the decision.