Use AI as a Research Assistant
This content was developed with AI assistance and is regularly reviewed for accuracy.
Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.
This playbook shows you how to use an assistant's research mode to gather and organize information on any topic, and then do the one thing that separates usable research from confident nonsense: check the citations. You will finish with a summary where every claim carries a label saying how much you trust it.
Last tested with Claude Research (Pro plan), Gemini Deep Research, and Perplexity in October 2026.
What you'll accomplish
- Write a research brief with scope, preferred sources, and a verification plan
- Run it in a research mode that plans searches, reads pages, and cites them
- Check citations with a three-question test
- Ask for the strongest counter-case
- Produce a summary with a confidence label on each claim
Prerequisites
- An assistant with web search. Research modes that read dozens of sources and write a cited report are mostly paid features: Claude Research is on Pro and above, Gemini's Deep Research is in the Gemini app and was upgraded in April 2026, and ChatGPT's deep research has usage limits that depend on your plan. Perplexity gives cited search on a free plan with a daily cap. Claude's free plan includes web search but not Research; it still works for Steps 3 to 5.
- A topic you genuinely need to decide something about
- 30-45 minutes, most of it waiting and checking
The playbook
Step 1: Write the research brief
Goal: Turn a vague interest into a question with scope, preferred sources, and a definition of what counts as evidence.
Action: Use the six-field brief from Your First Delegated Task, with extra care on resources and verification; those fields are what make research checkable.
Objective: Compare the five-year total cost of owning a mid-range electric sedan versus a comparable gas sedan in the US.
Context: I'm deciding whether my next car is electric. I drive about 12,000 miles a year and can charge at home.
Resources: Prefer primary sources: manufacturer pricing pages, government data (fueleconomy.gov, IRS incentive pages, EIA electricity prices), and insurer or consumer-group studies. Cite every number.
Constraints: US market, cars available new in 2026, sources from the last 12 months where possible. Include purchase price, fuel or charging, maintenance, insurance, and incentives.
Output: A comparison table, a short narrative of what drives the difference, and a numbered source list.
Verification: Next to each figure, say whether it came from a primary source, a secondary summary, or your own estimate, and give the publication date of the source.
Checkpoint: Your brief names the kinds of sources you trust and asks for dates and labels.
Step 2: Choose the mode and run it
Goal: Match the depth of the tool to the depth of the question.
Action: For a quick factual question, the normal chat with search on is enough. For a question like the one above, switch to the research mode (the "Research" or "Deep Research" toggle near the message box). The assistant will write a plan, run many searches, read the pages, and return a cited report. Expect five to twenty minutes; start it and come back. Gemini's Deep Research can also read your Gmail and Drive if you grant permission, which helps when part of the evidence is in your own files.
One piece of 2023 advice to drop: you no longer need to paste articles in because "the AI's knowledge has a cutoff." Current models were trained on data from within the last few months and search the web by default. The question is no longer whether it can know; it is whether it looked, and where.
Checkpoint: You have a report with numbered citations and a sources list, saved in the conversation.
Step 3: Check the citations
Goal: Confirm the facts that would change your decision.
Action: This step is the whole reason the playbook exists. In 2026, assistants hallucinate rarely on short, grounded tasks, but error rates climb with long documents, with legal and medical specifics, and above all with citations. By September 2026, trackers counted roughly 2,000 court cases involving fabricated AI citations, up from about 200 a year earlier (reported). A 2026 study of invented scientific references called its own figures a lower bound. Professionals keep failing this check, not just beginners.
Pick the three to five claims that would change your conclusion and run each through three questions:
- Does the source exist? Click the link. A dead link or a page about something else is a failed check.
- Does it say that? Find the number or quote on the page. A newer failure pattern is a real source with an invented quotation, so a working link is not enough.
- Is it current and authoritative? Check the date and who published it. A 2022 blog post about EV incentives is not evidence for 2026.
Ask the assistant to help you do this faster:
For claims 2, 5, and 7 in your report, quote the exact sentence from the source that supports the figure, with the URL. If you cannot find a sentence that says it, tell me so and downgrade the claim to an estimate.
Anything that fails a question gets corrected or marked unverified. Do not let it stay in the report with a footnote attached.
Checkpoint: Every load-bearing claim has passed all three questions or has been relabeled.
Step 4: Ask for the other side
Goal: Get a balanced view before you decide.
Action: Research reports tend to answer the question as asked. Push back explicitly.
Now make the strongest case against buying an electric sedan in 2026, using the same source standards. Include practical downsides that EV advocates tend to play down, and cite each one.
Run the same three-question check on the two or three claims that most weaken your original conclusion.
Checkpoint: You have both the supporting and the opposing case, each with checked sources.
Step 5: Compile with confidence labels
Goal: Produce a summary you can share without a disclaimer.
Action: Ask the assistant to consolidate the conversation and to label each claim.
Based on our full conversation, write a summary with these sections:
1. Key findings (bullet points, each ending with a label: Verified by me / Source checked, not independently confirmed / Unverified)
2. Five-year cost comparison table
3. Main advantages and disadvantages
4. Open questions I should still look into
Keep it under 500 words and keep the source list.
Copy the summary into your own document. If you will keep working on this topic, put it in a project so later questions start with this context instead of from scratch; the Fluency page Supply Context shows how.
Checkpoint: A verified summary is saved outside the app, with labels your reader can see.
Common pitfalls
- Trusting a report because it has footnotes: Footnotes are claims about sources, not proof. Click them.
- Checking the easy claims: Verify the numbers that would change your decision, not the ones that are easy to find.
- Accepting the first framing: Ask what you are missing and what the other side would say.
- Using research mode for everything: It is slow and uses your plan's limits. A quick question with search on is often enough.
- Filing or publishing AI references unchecked: This is the mistake the court sanctions keep punishing. Never let a citation leave your hands without the three-question check.
Key takeaways
- A research brief names the sources you trust and asks for dates and labels
- Research modes plan, search, read, and cite; your job moves from gathering to checking
- Three questions per citation: does it exist, does it say that, is it current and authoritative
- Always ask for the strongest counter-case and check it to the same standard
- Label every claim in the summary so readers know what you verified
Next steps
Continue to: Draft Better Emails and Documents with AI to apply the same brief-and-check habit to everyday writing.
Continue to: Check: Verify, Iterate, Restart for the full set of verification habits, including how to check what an agent did, not just what it said.
Continue to: Decide: What to Delegate, Which Tool, Which Mode for choosing between fast, thinking, research, and agent modes.