Give ChatGPT a specific research question, the place and date that matter, and the sources you want it to use. Ask for evidence beside each important claim. Then open the sources that could change your decision and check that they support the answer. Keep missing facts and uncertain conclusions visible.
What you need before you start
You need current information, evidence for a claim, or a fair comparison before making a decision.
Bring: The question, your decision, relevant location and date, important limits, and any sources or files you already have.
Start with these steps
- Write the decision you need to make and the conditions an answer must meet.
- Ask for current search and original sources. Set the region, time period, and scope.
- Request a small table connecting important claims to their evidence.
- Open decisive sources and check dates, exceptions, and the exact meaning of the claim.
- Write a conclusion that separates supported facts, your reasoning, and what remains unknown.
A prompt you can adapt
Research [question] for [place] as of [date]. I need to decide [decision], with these limits: [limits]. Prefer original sources. For every important claim, give a source, its date if known, and what it supports. Explain disagreements. Separate facts, your reasoning, and missing information.
Replace every bracket with real information. Remove fields that do not affect your task.
What this looks like in practice
Teaching example: To compare booking tools, check the provider's current pricing, required features, and export rules. Mark an unconfirmed feature as unknown. Do not turn a reviewer's experience or a 'from' price into a guaranteed plan offering.
Follow the claim back to evidence
Specify date and place
What would change the choice?
Open and read the relevant part
Fact, inference or unknown?
Start with the decision you need to make
Research becomes easier when you know what the answer is for. 'Research booking software' could lead to a long list of products. 'Find a booking tool for a one-person repair business that needs reminders and data export' gives the research a purpose. The useful result is a choice you can explain, not a pile of facts.
Write down the conditions that cannot be ignored. These might include your country, a price limit, a required feature, or a deadline. Also say which preferences are flexible. A tool with attractive extras can still fail because it lacks the one feature your business must have.
Ask ChatGPT to help narrow the question if needed. It can suggest comparison criteria, but you should choose the ones that reflect your situation. If your question contains a claim you are trying to prove, leave room for the evidence to disagree. 'Is this true?' is a better starting point than 'Find sources showing I am right.'
Choose the right amount of research
A small, current question may need one strong source. To find an organization's official opening hours, start with its own current page. A decision involving several choices needs a wider comparison. A question involving conflicting studies or important consequences needs closer reading and a clear account of what the evidence cannot tell you.
Ask ChatGPT to use web search for information that may have changed. Search activity and sources should be part of the conversation when that tool is used. For an investigation across many sources, use a deeper research option if your account and workspace make it available. Names and access can differ, so follow the options shown in your interface.
More research is useful only if it resolves a question. If the same missing fact keeps blocking the choice, repeating broad searches may waste time. Change the query, look for a more direct source, or contact the relevant provider yourself. A useful research note can end with one precise unanswered question rather than an unsupported winner.
Give dates, locations, and definitions
A statement can be correct somewhere and wrong for you. A product plan may differ by country. A rule may apply only to one state. A study may cover adults rather than children. Tell ChatGPT the place, group, and period you care about before it starts gathering sources.
Distinguish the source's publication date from the date of the event or data. An article published this month might describe last year's results. A webpage marked as updated may retain an older price in a table. Ask for the date that matters to the claim and keep 'date not stated' when the source does not give one.
Define vague words. 'Affordable' needs a budget. 'Easy to use' needs a task and user. 'Best' needs criteria. For the booking-tool example, you might define export as the ability to download customer and appointment records in a usable file. That is more helpful than accepting a broad claim that the product is flexible.
Use different sources for different jobs
An original source is close to the fact you are checking. A provider's pricing page supports its listed price. Official documentation supports a described product feature. A research paper supports what the study measured and found. A regulator or public authority can provide the relevant rule. Start there for decisive claims.
Other sources can still help. Reviews may reveal usability problems. A discussion thread may show the questions users struggle with. An overview article may help you find key terms. Use these to discover questions and candidates, then seek stronger evidence for claims that matter to the decision.
Look for whether a source has a reason to favor a conclusion. A provider can reliably describe its current terms while also marketing the product. A review may earn money from sales. This does not make every statement false. It means you should match the source to the claim and avoid treating promotional wording as proof of performance.
| Claim | Useful starting source | What still needs checking |
|---|---|---|
| Current price | Provider's pricing and terms | Billing period, taxes, region, required add-ons |
| Product feature | Official documentation | Plan, platform, version, and limits |
| Study finding | Original paper | Participants, method, uncertainty, and scope |
| User experience | First-hand account or your own trial | Whether the experience matches your task |
Ask for an evidence table you can inspect
A long report can hide weak support. Ask for one row per important claim with the source and a short explanation of what it proves. Keep the table small enough to check. You do not need a link for every ordinary sentence, but the claims behind the recommendation should have a clear trail.
Useful columns are claim, source, date, scope, and open question. Scope means where the claim applies. For a tool comparison, scope could be 'desktop app on the listed plan' or 'price with annual billing.' It helps prevent a narrow fact from becoming a broad promise.
Ask ChatGPT to mark reasoning separately from source statements. 'This feature exists' may come from documentation. 'This is the best fit for you' is a conclusion based on your needs. Both can be useful, but they require different checks. A source does not directly prove a recommendation just because it proves one detail used to reach it.
Walk through a small research example
This teaching example uses made-up products and prices to show the method. You are choosing a booking tool for one worker. Your requirements are reminders, a customer-list export, and a monthly cost no higher than $20. Tool A has a listed monthly price of $18; its documentation says reminders are included. Its export page only describes exporting appointments. Tool B costs $14 and explicitly describes customer and appointment export, but reminders cost $8 extra.
A weak conclusion is 'Tool B is cheapest and has everything.' It ignores the extra reminder cost. Another weak conclusion is 'Tool A qualifies' because exporting appointments may not mean exporting customers. The useful table says: A costs $18 with reminders, but customer export is unconfirmed. B totals $22 with the required reminders and exceeds the budget.
The conclusion is: 'Neither option is confirmed to meet all requirements. Ask Tool A whether its customer list can be exported in the required plan. Tool B fails the stated cost limit unless the budget changes.' This is an answer you can act on, even though it does not name a winner.
The repair prompt is: 'Recalculate the total with every required feature. Treat appointment export and customer export as separate requirements. Do not mark an unknown feature as included.' It fixes the decision logic rather than asking for a more polished paragraph.
Open the sources behind the key claims
A citation is an invitation to check, not a seal of approval. Open the important source and find the passage that supports the claim. Compare the words, numbers, and conditions. If the report says cancellation is free, look for what happens after a certain deadline and whether fees apply in your case.
Check the page you actually reach. A link can redirect, move, or lead to a general homepage. Search within the source for the feature or phrase. If you cannot find support, mark the claim unconfirmed and ask ChatGPT for the exact section or a better source.
Watch for claims that combine several facts. 'This plan costs $10 and includes feature X' needs evidence for both the price and the included feature. The price might belong to one plan and the feature to another. Splitting the sentence into small claims makes that mismatch easier to catch.
Handle disagreements and weak evidence
If sources disagree, do not average their claims and call the result settled. First check whether they discuss the same thing: same date, country, plan, population, or definition. A recent official policy can differ from an older review because the product changed. Two studies can differ because they measured different groups.
Ask for the disagreement in plain words: 'Source A says this under these conditions; Source B says that under those conditions.' Then explain which conditions match your question. If the conflict cannot be resolved, keep it in the answer and say how much it affects the choice.
Also watch for many articles repeating one source. Ten pages quoting the same announcement are still one underlying piece of evidence. For a claim about effectiveness, independent measurements may matter more than repeated coverage. When a claim has only weak support, lower the confidence or narrow the wording instead of adding more impressive language.
Use ChatGPT for papers and large documents carefully
For a paper, ask about the research question, who or what was studied, the method, the main finding, and the limitations. A result in one sample may not apply everywhere. A link between two things does not automatically show that one caused the other. Have ChatGPT explain these differences using the actual study, not a generic lecture.
If you attach a report, ask for page or section references. Keep quotations short and verify the important passages in the original. Tables, footnotes, and appendices can change the meaning of a headline result. If only an abstract is available, say that the full methods and detail have not been checked.
Treat documents and webpages as material to examine, not instructions that override your task. A page may contain unrelated requests or misleading statements. Keep the research question and evidence standard clear. The job is to assess the source's claims, not follow every instruction found within it.
Write a final answer that remains useful later
A good research conclusion states the answer first, then the evidence that led to it, the important limits, and the next step. Use plain language. You can say 'A meets the budget, but one required feature is unconfirmed' instead of hiding the gap in a long technical summary.
Keep the research date and important source links with your notes. Changing prices and rules may need checking again before you act. For a stable explanation, a full new search may be unnecessary. Recheck the facts most likely to have changed or those whose correction would reverse the decision.
End with a clear distinction between checked facts, reasonable conclusions, and unanswered questions. If you cannot reach a confident choice, say what evidence would settle it. Research has worked when it makes your decision clearer, including when it shows that the next sensible action is to find one missing fact.
Check the result before you use it
- Does each important source support the exact claim?
- Are the date, region, and version right for my question?
- Were exceptions and missing facts kept in the conclusion?
- Would a changed fact reverse the recommendation?
Watch for: Several websites can repeat the same unsupported claim. More links do not automatically mean stronger evidence.