How to research with ChatGPT and Claude

Choose between a quick search and a research task

A question such as “Where is this vendor’s current help page?” needs a small lookup. Choosing between several suppliers needs a structured comparison across sources. The amount of research should match the decision.

OpenAI’s web-search documentation describes search with citations and research capabilities within available tools. Anthropic documents web search and a separate Research feature. Access depends on the product, account and workplace settings; Claude Research requires a paid plan and web search enabled.

An ordinary chat answer is not evidence that live search happened. Ask for current sources and check that the result actually includes relevant links.

Write the decision before asking for research

A broad request such as “research scheduling software” can return a long list that helps little. Define who will use the result, the constraints and what you want to decide.

For a fictional Australian service business, the decision might be which three scheduling tools deserve a demonstration. Useful constraints could include mobile access, job assignment, exportable records and Australian support hours. State which are essential and which would be convenient.

This guide uses a proposed workflow, not a tested ranking of vendors. Your own requirements and available evidence should determine the shortlist.

Build the brief, then inspect the sources

1. Write the question, geography, date sensitivity and required output. Give a short list of essential criteria.
2. Ask the tool to outline its approach before a broad investigation. Remove unrelated questions.
3. Prefer original sources: the vendor’s documentation for features, its pricing page for published prices and official notices for changes.
4. Request a source beside each important claim and a visible “not confirmed” for missing evidence.
5. Open the sources that determine the shortlist. Check the date, plan, region and qualifications.
6. Write a recommendation that separates verified facts, reasonable inferences and questions for the vendor.

Copy this original research prompt

Replace the example requirements with your actual approved brief. Ask for a small shortlist so you can inspect the evidence.

Research three scheduling tools that may suit a fictional Australian service business with 15 staff. Essential: job assignment, mobile access and exportable job records. Useful: support hours that overlap Sydney business hours. Use current official product documentation and pricing pages where available. For each criterion give the finding, source URL and any plan or regional restriction. Do not infer a feature from marketing language. Mark missing evidence “not confirmed”. Separate published facts from your assessment. End with questions for a product demonstration; do not purchase or contact anyone.

A citation is the start of a check

Open the cited page and read the passage behind the claim. The page might describe a different plan, an old version or an announced feature that has not reached your account. A correct link can still be used to support the wrong conclusion.

Ask when the underlying event occurred as well as when the article was published. A newly updated page may describe an older launch. Conversely, a release note can describe a staged rollout that is still unavailable to some users.

For a contested claim, look for genuinely separate evidence. Several pages repeating the same press release are not several independent confirmations. If the evidence is thin, narrow the conclusion.

Keep facts, inference and missing information visible

Use three labels in your final notes. Fact: the official documentation lists CSV export on a named plan. Inference: that may make migration easier. Open question: whether the export includes the fields your team needs.

Do not turn an inference into a fact just because it seems likely. A product offering an integration does not prove it supports your exact action. A vendor saying it serves Australia does not establish its support hours.

Check calculations separately. If you compare annual costs, use consistent currencies, billing periods and user counts. Record which inputs came from published prices and which remain assumptions. Ask your accountant about any tax treatment that affects the decision.

Use research tools without delegating the decision

Connected company sources may be useful when authorised, but external web evidence and internal notes should remain distinguishable. Check what access you are granting and whether your workplace allows the connection.

End with a decision-ready result: shortlist, reasons, unresolved questions and the next verification step. Save the source URLs and the date you checked them. For important purchases, a demonstration using your own sample workflow can resolve gaps that a research report cannot.

For help teaching this research method to your team, use the workshop enquiry form on the 27 AI homepage.

Common questions

Does deep research guarantee a correct answer?

No. A broader search can collect more evidence, but the sources and reasoning still need review. Check the claims that drive your decision.

Why is Research missing from my account?

Availability can depend on plan, rollout, platform and workplace settings. Claude’s Research documentation also requires web search to be enabled. Check the current official help page for your product.

Can I use an AI answer without citations?

For a task based only on text you provided, check it against that text. For current external claims, ask for sources and verify them before relying on the answer.