To use AI for research well, start with a question you can investigate and finish with evidence you can trace. The assistant can help you find directions, organize documents and compare arguments. You still need to inspect the sources behind the answer.
This guide gives students, writers and small teams a repeatable process. It works for an article, a product investigation or a background report, and it does not depend on one particular chatbot.
Step 1: turn a topic into a research question
“AI in education” is a topic. “What evidence supports using AI feedback for first-year writing students?” is a question. The second version tells you which evidence matters and which information belongs outside the project.
Before searching, write down your audience, time period, geography and intended output. A 600-word explainer needs a different research plan from a dissertation chapter. These limits keep a broad conversation from expanding indefinitely.
Help me narrow this topic into three research questions. For each, identify the population, the outcome being investigated and the kinds of primary sources I would need. Do not answer the questions yet.
Step 2: choose a tool for the job
A tool with web search is useful for discovering current information. A document-focused assistant is useful when you already have a collection of papers. A spreadsheet or reference manager helps you retain the evidence after the chat is over.
For example, Perplexity’s Pro Search documentation describes linked sources and follow-up searches. Google’s notebook source guidance explains how imported documents provide context for questions. Those features serve different stages of research.
Choose based on the task rather than the longest feature list. If you need to compare three supplied reports, uploading those reports to an approved document tool may be more useful than asking a general assistant to search the entire web.
Step 3: request sources before conclusions
Ask for a short source shortlist with publication dates and a reason each item matters. Prefer the original study for research findings, official documentation for product behavior and a directly published report for organization statistics.
Find primary sources about this question published within the last three years. For each source, give the title, publisher, publication date, link and one sentence explaining its relevance. Separate original research from opinion. Say when no suitable primary source is available.
Open the links yourself. Confirm the source exists, the title matches and the date is relevant. A citation-shaped string is not evidence. If a link fails, search for the exact title through the publisher rather than accepting a substitute silently.
Step 4: build a small evidence table
Use one row per important claim. An evidence table makes gaps visible before they become confident paragraphs.
| Field | What to record |
|---|---|
| Claim | The precise statement you may use |
| Source | Title, publisher, date and link |
| Evidence location | Page, section, table or passage |
| Scope | Who, where and when the evidence covers |
| Limitation | What the evidence cannot establish |
Suppose a survey reports that some employees prefer a new tool. That supports a statement about reported preferences in the surveyed group. It does not automatically establish higher productivity, better work quality or a benefit for every industry.
A reusable AI research prompt template
Replace the brackets before running this prompt. It separates discovery from evidence checking so that a polished answer does not hide missing research.
My research question is [specific question]. The audience is [audience], and the scope is [dates, place and exclusions]. First propose a search plan and up to five relevant primary sources. For each, give its title, publisher, date and URL. Do not invent a source or claim to have opened a page you cannot access. After I supply verified sources, build a claim-evidence table with a page or section reference and limitations. Separate direct findings, your interpretation and unanswered questions. Do not draft conclusions until I ask.
Use a web-enabled tool for the discovery step. When working from supplied PDFs, ask for answers limited to those files and check the cited passages. Our NotebookLM document tutorial covers that workflow; the source-backed blog research guide shows how to move verified notes into an article.
Worked example: compare software without inventing evidence
Suppose your question is, “Which of three file-sharing services supports the client permissions I need?” Define the required permissions first. Then open each provider’s current sharing documentation and record the account or plan it describes.
| Evidence-table field | Example entry to complete |
|---|---|
| Requirement | Only named client recipients should be able to open the delivery. |
| Source location | Provider documentation URL, section title and access date. |
| Supported finding | The exact documented permission and the account or plan it applies to. |
| Unresolved check | Whether the control appears in our own account; test with a non-sensitive sample. |
This is a research template, not a claim that all services offer the same controls. If the document does not mention a feature, record “not established from this source,” rather than “not supported.” A research answer is ready to use when every material claim has a checked source and any remaining uncertainty is visible.
Step 5: summarize with boundaries
When you provide a paper, ask the assistant to distinguish the author’s conclusions from its own interpretation. Keep your summary focused on the research question rather than requesting every detail.
Using only the attached report, extract the main finding relevant to my question, the method, the sample and the limitations. Give a page or section reference for each finding where available. If something is missing, write “Not reported.”
Check the passages behind the extraction. For scanned documents, first confirm that the text was recognized correctly. A misplaced decimal point or a table heading read as a result can undermine the entire summary.
Step 6: look for disagreement
Ask what would weaken your current explanation. A useful research process includes conflicting evidence, different definitions and limits on generalization.
Compare these sources on the same question. Identify where they agree, where they disagree and whether the differences may come from sample, date, method or definition. Keep observations separate from suggested explanations.
Do not count three articles repeating one report as three independent findings. Follow the references back to the original evidence and record how many genuinely separate sources you have.
Step 7: write from your verified notes
Build your outline from the evidence table. Put supported claims in the main argument, limitations beside the relevant claims and unresolved questions in a separate section. Write the final explanation in your own voice.
A final review should ask whether each number has a source, whether comparisons use compatible definitions and whether the conclusion is narrower than the evidence. Students should also follow their institution’s rules for AI assistance and disclosure.
Protect your source material
Use public or approved documents for practice. Remove personal identifiers and confidential material unless your organization has authorized that tool and use case. Review sharing settings before inviting others into a research workspace. Our guide to AI workflows and client privacy explains a practical way to separate reusable instructions from private inputs.
Frequently asked questions
Can AI do my literature review for me?
It can assist with discovery, extraction and organization. You must verify coverage, read the relevant work and make the judgments that your review requires.
Are linked citations automatically reliable?
No. Check whether the linked source actually supports the nearby statement and whether it is appropriate for the question.
What should I do if sources conflict?
Compare their methods, definitions and scope. Report the disagreement instead of forcing a single answer that the evidence does not support.
Start with one question and three sources. A small, traceable research file is more useful than a long answer whose evidence you cannot locate.
Continue your AI productivity workflow
This guide is part of our AI Productivity Tools & Workflows pillar guide. Use it to connect research, drafting and review into one repeatable process.
