Beforemark Guided Journey
Step 2 of 3: Learn the Research Method
Follow a repeatable process for asking better questions, inspecting sources, and forming careful conclusions.
Learn a Research Method You Can Use With Any Major AI Tool
AI can make research faster, but speed is useful only when the result deserves your trust. This tutorial teaches a simple method for using AI to explore a topic without handing over your judgment.
You will learn to:
- turn a broad interest into a clear research question;
- break the question into manageable parts;
- ask AI to find evidence instead of merely producing an answer;
- inspect sources and verify important claims;
- recognize conflicts, gaps, and uncertainty; and
- write a conclusion you can explain in your own words.
Before starting: Choose and set up a tool on the AI Research Journey Hub. If you already have a web-connected AI tool and can open its citations, continue below.
Where you are: Tool selected and set up → Learn the method → Complete the First Project
What AI Can—and Cannot—Do for Research
AI is good at helping you:
- identify useful questions;
- find possible sources;
- summarize complicated material;
- compare ideas or viewpoints;
- organize findings; and
- reveal areas that need more investigation.
AI is not reliable enough to:
- guarantee that every statement is accurate;
- decide whether a source is trustworthy for you;
- ensure a citation supports the sentence beside it;
- know every missing fact or perspective;
- replace subject-matter expertise; or
- make high-stakes decisions on your behalf.
The right mental model is simple: AI helps you investigate; evidence supports the conclusion.
The Beforemark AI Research Method
Step 1: Define what you are trying to understand
Begin with the outcome, not a giant topic.
Broad topic:
Home solar panels
Useful outcome:
I want to understand whether home solar panels are likely to make financial sense for a typical homeowner in North Texas who expects to stay in the house for ten years.
The second version gives the research boundaries. It tells you which information matters and which information can wait.
Write one sentence:
I want to understand __________ so that I can __________.
If the research involves a decision, name the decision. If it is purely educational, name what you want to be able to explain afterward.
Step 2: Turn the goal into a clear research question
A useful research question is:
- specific enough to investigate;
- neutral rather than designed to confirm what you already believe;
- bounded by relevant place, time, audience, or conditions; and
- answerable with evidence.
Weak question:
Are solar panels worth it?
Stronger question:
Under what conditions do residential solar panels provide a reasonable financial return for a North Texas homeowner planning to remain in the home for ten years?
Use this prompt to improve your question:
I want to understand [topic] so I can [outcome]. Help me turn this into one neutral, specific research question. Identify any important boundaries I should add, but do not answer the question yet.
Check the suggested question yourself. Remove assumptions, loaded wording, or unnecessary complexity.
Step 3: Break the question into subquestions
One large question usually hides several smaller questions. Subquestions reduce the chance that the AI will give you a polished but shallow answer.
For the solar example, useful subquestions might include:
- What are the typical upfront and ongoing costs?
- How much electricity might the system produce under local conditions?
- Which incentives or policies currently apply?
- How do financing terms affect the return?
- What happens if the homeowner moves early?
- What maintenance, warranty, roof, or insurance issues matter?
- Which assumptions have the greatest effect on the conclusion?
Prompt:
Break this research question into five to seven neutral subquestions. Include costs, benefits, risks, alternatives, important assumptions, and evidence that could change the conclusion. Do not research them yet.
Keep only the questions needed for your outcome. More questions do not automatically produce better research.
Step 4: Set the evidence standard before searching
Tell the AI what kinds of sources you want. Without direction, it may favor sources that are popular, easy to summarize, or optimized for search rather than sources best suited to the claim.
Use this source order when practical:
- Primary and authoritative sources: government agencies, laws and regulations, official statistics, original research, standards bodies, official product documentation, and direct organizational records.
- Strong secondary sources: systematic reviews, established academic or professional institutions, and reputable reporting that links to original evidence.
- Experience and commentary: expert analysis, industry publications, forums, and user experiences—useful for context, not as sole support for major factual claims.
Prompt:
Research the subquestions using current sources. Prioritize primary and authoritative sources. Use strong independent secondary sources where they add context. Separate factual evidence from opinion or anecdote. Cite every important factual claim with a direct link.
For time-sensitive topics, add an exact recency boundary:
Use sources current as of [month and year], and identify any information that may change soon.
Step 5: Run the first research pass
Submit the research question, subquestions, boundaries, and evidence standard together. Ask for a structured response rather than an instant verdict.
Copy-ready prompt:
Help me research this question: [research question]
My purpose is: [what I want to understand or decide]
Investigate these subquestions:
- [Subquestion]
- [Subquestion]
- [Subquestion]
- [Subquestion]
- [Subquestion]
Prioritize primary and authoritative sources, then use strong independent secondary sources for context. Cite each important factual claim with a direct link. Distinguish facts, reasonable inferences, and unknowns. Identify conflicting evidence, important missing information, and assumptions that could change the conclusion. Do not give a final recommendation yet.
Do not worry if the first response is incomplete. It is a map for the next research pass, not the finished result.
Step 6: Inspect the sources behind the answer
Open the citations supporting the claims that matter most. Check:
- Relevance: Does the source address the actual claim?
- Authority: Who created it, and are they qualified or responsible for the information?
- Evidence: Does it show data, methods, documentation, or merely make an assertion?
- Date: Is it current enough for the question?
- Context: Did the AI leave out qualifications, limitations, or conditions?
- Interest: Does the publisher benefit from persuading you?
When the exact supporting passage is hard to locate, ask:
For the claim [paste claim], identify the exact source, the relevant section or passage, and explain how the source supports the claim. If the source does not directly support it, say so and find a better source.
Then open the source yourself. Never rely on the AI’s assurance that it checked correctly.
Step 7: Improve source diversity
Five links do not necessarily mean five independent sources. Multiple articles may repeat the same press release, study, dataset, or unsupported claim.
Look for diversity across:
- source type;
- publisher or institution;
- underlying data;
- relevant viewpoints;
- geography or population, when relevant; and
- publication date.
Prompt:
Review the sources used so far. Identify which ones depend on the same original evidence or repeat one another. Then find independent sources that could confirm, qualify, or challenge the key claims.
Source diversity does not mean treating every opinion as equally credible. Weight evidence by quality and relevance.
Step 8: Search for conflicts and missing information
Weak research asks, “What supports this answer?” Stronger research also asks, “What would make this answer wrong or incomplete?”
Prompt:
Act as a critical research reviewer. What credible evidence conflicts with the current findings? Which affected groups, limitations, alternatives, or important variables may be missing? Identify what additional evidence would most change the conclusion.
When sources disagree, determine whether the disagreement comes from:
- different definitions;
- different populations or conditions;
- different time periods;
- different research methods;
- commercial or ideological incentives;
- normal uncertainty; or
- genuinely contradictory evidence.
Do not ask the AI to “pick the winner” before you understand why the sources differ.
Step 9: Build an evidence table
A short evidence table prevents claims and citations from becoming separated.
| Key claim | Best supporting source | What the source actually shows | Limitations or conflicts | Confidence |
|---|---|---|---|---|
| [Claim] | [Direct link] | [Your note] | [Caveat] | High / Medium / Low |
Write the “what it actually shows” column after opening the source. Keep this table focused on the claims needed for your conclusion.
Step 10: Write a careful conclusion in your own words
Before asking AI for polished prose, explain the findings yourself:
- What is the best-supported answer?
- What evidence matters most?
- What remains uncertain?
- Under which conditions would the answer change?
- What should the reader not conclude?
Then use AI as an editor:
Using only the verified claims and notes below, help me organize a concise conclusion. Preserve all uncertainties and limitations. Do not add facts, sources, or confidence that are not present in my notes. Clearly separate confirmed findings, reasonable inferences, and unresolved questions.
If you cannot explain the conclusion without reading the AI response, you do not yet understand it well enough to use it.
Tool-Specific Notes
The research method stays the same across tools. Only the controls and strengths differ.
Perplexity
- Open the numbered citations while reading.
- Use follow-up questions to challenge a claim or request primary sources.
- Normal Search is enough for this tutorial; advanced research modes are optional.
ChatGPT
- Ask it to search the web when current facts or sources matter.
- Open the inline citations and Sources panel where available.
- Tell it not to answer from memory when the question requires current verification.
Gemini
- Normal web-connected use is sufficient.
- If Deep Research is available, review and edit its research plan before starting the report.
- Treat connected Google material as one source category; still seek independent evidence when needed.
Claude
- Turn on web search for current information and citations.
- Uploaded documents can be the starting evidence, but they may not represent the complete picture.
- Research mode is optional; the normal method still works without it.
Essential Safeguards
Accuracy
Verify the claims that materially affect your conclusion. The more costly, consequential, surprising, or precise a claim is, the stronger the evidence should be.
Citations
Do not cite an AI tool as the source of a fact when an original source is available. Cite the original material you opened and reviewed.
Privacy
Do not upload confidential, personal, proprietary, or restricted information unless you are authorized to do so and understand the tool’s data controls. Redact or generalize whenever possible.
High-stakes use
For medical, legal, financial, employment, safety, or similarly consequential matters, confirm findings through authoritative sources and a qualified professional. AI research can prepare better questions; it cannot assume professional responsibility.
Are You Ready for the First Project?
Before continuing, confirm:
You now know the method. The next page turns it into a complete task and gives you a standard for deciding whether the result is good enough.
Previous: Choose or change your
tool
Next: First AI Research Project