What you will learn
- Treat a model’s response as a starting point for investigation.
- A citation must support the particular claim beside it.
- Keep missing evidence and unresolved questions visible.
1. Make the research question answerable
A broad question such as “What is the future of this industry?” often produces a general overview. Define the topic, time period and scope. Ask about a particular feature change, a measured trend or the difference between two methods for a specific use case.
Before searching, write down the decision this research should inform. A first map of the topic may be enough for orientation. Publishing a figure requires more: you need its definition, date and measurement method, as well as the reported value.
2. Record evidence beside each claim
Use an AI tool to suggest topics and possible lines of inquiry, then open the sources yourself. For a product feature, read the maker’s documentation. For a statistic, find the original report and definition. A related link is not automatically evidence for the sentence next to it.
Keep three columns in your research notes: the claim, the source and supporting passage, and the review status. This makes it clear which statements have been checked and which are still suggestions, assumptions or interpretations.
Research question: [your question] List claims that can be checked. For each claim, identify an accessible source and the relevant section. Separate reported facts from your interpretation. Say explicitly if you have not opened a source or lack sufficient evidence. Do not give a firm conclusion before the sources have been checked.
3. Check dates, definitions and scope
A source that was accurate in the past may not answer today’s question. Check publication and update dates, product names, versions and the population being studied. Keep the distinction between an experimental sample and a general finding in view.
When two sources disagree, first check whether they are measuring the same thing. Units, time periods and metric definitions may differ. Placing two figures side by side without those explanations can create a misleading comparison.
4. Preserve uncertainty in the result
Some questions do not have a clear answer in the available information. Keep a “What we still do not know” section rather than filling the gaps with guesses. Record missing data, conflicting sources and what would be needed to continue the investigation.
If generated text sounds unusually certain, ask the model to list its assumptions. Its own second answer is still not independent evidence. The main verification must return to external data and sources that you can inspect.
5. Write the final synthesis yourself
After checking the evidence, retain supported claims in the final text. Compare quotations with their originals and place each citation near the claim it supports. Help the reader distinguish what was observed from your interpretation of it.
Bring a knowledgeable reviewer into the process for specialist topics or consequential decisions. A readable response with several links can help organize the work, but it does not replace the researcher’s responsibility or appropriate expert review.
Your next-step checklist
Review each item before using the result.
Further reading
This guide suggests a workflow. Check the official source and the specific product page for current features and terms.
Related tools in the ElmosGPT catalog
Perplexity
NotebookLM
ChatGPT