A repeatable system for researching anything with an AI research mode (Perplexity, ChatGPT or Gemini), from framing the first question to a decision memo you can act on.
A repeatable system for researching anything with an AI research mode (Perplexity, ChatGPT or Gemini), from framing the first question to a decision memo you can act on.
A research mode that cites its sources is the first thing I open for anything I don't already know cold: a competitor's pricing, a regulation I'm not sure about, a market I'm about to pitch into. Perplexity, ChatGPT's research modes and Gemini's research modes all work this way, and this workflow runs the same in any of them. Not because they're magic. Because they cite their sources, and citations are what turn an AI answer into something you can trust enough to act on.
The problem isn't the tool. It's that most people use it like a search engine and expect research-grade output. One question, one answer, done. That produces a confident paragraph built on two blog posts and a Reddit thread. This is the workflow I actually run: how I frame the first question, how I check what comes back, the follow-up questions I always ask, and how I turn the whole thread into a one-page memo before I act on it.
Two minutes of setup before your first question saves you from redoing the whole thread later.
A search engine wants three keywords. A research assistant wants a brief: who you are, what decision this feeds, what you already know, and what shape you want the answer in. The single biggest quality jump in this whole workflow comes from this one habit.
I'm researching [TOPIC/DECISION] for [YOUR SITUATION: business type, size, or context].
What I already know:
[WHAT YOU KNOW SO FAR, even if it's rough]
What I need from this research:
[THE ACTUAL DECISION THIS FEEDS, e.g. "pick one vendor," "decide whether to expand," "understand if this regulation applies to me"]
Constraints:
- Budget: [NUMBER OR RANGE, or "n/a"]
- Timeline: [WHEN YOU NEED TO DECIDE]
- [ANY OTHER HARD CONSTRAINT]
Give me a direct answer with sources, not a generic overview. If the honest answer is "it depends," tell me what it depends on.Citations are the whole reason a research mode beats a plain chat AI for research, but a citation isn't automatically a good source. Run this check on any answer that's going to inform a real decision. It takes under two minutes.
Before I act on this, stress-test it for me. For the claim that [SPECIFIC CLAIM FROM THE ANSWER]:
1. Which of your sources are primary (the company/organization itself, an official filing, original reporting) versus secondary (blogs, aggregators, forum posts)?
2. Are any of your sources more than 12 months old? Flag them.
3. Is there a more recent or more authoritative source that contradicts this?
4. On a scale of "well-established fact" to "single source, take with caution," how confident should I be in this specific claim?The first answer to any real research question is the surface. It's accurate as far as it goes, and it almost never goes far enough to actually decide something. I run every research thread that feeds a real decision through the same four rungs before I stop.
Run these as follow-ups in the same thread, in order:
1. Narrow this to my exact situation: [YOUR SPECIFIC CONSTRAINTS]. Does the answer change?
2. Put numbers on this wherever you can, actual costs, percentages, or timelines, not just qualitative descriptions.
3. Now argue the strongest case against this recommendation. What would someone who disagrees say, and are they right about anything?
4. What situation or constraint would make this recommendation wrong for someone? Do any of those apply to me?A long research thread is not a decision. It's raw material. Before you act, compress it into something you (or a partner, or a client) could read in 90 seconds and know exactly what to do. This is the step most people skip, and it's the one that actually makes the research useful instead of just interesting.
Turn this research thread into a one-page decision memo. Use this exact structure:
DECISION: [one sentence, what's being decided]
RECOMMENDATION: [one sentence, your call, stated plainly]
WHY: [3-4 bullets, the strongest evidence for the recommendation, with sources]
WHAT COULD GO WRONG: [2-3 bullets, the counter-case and edge cases from earlier in this thread]
NEXT STEP: [one concrete action, not "learn more" or "consider options"]
Keep it under 200 words total. No hedging language, no "it depends" without saying what it depends on. If the evidence doesn't support a clear recommendation, say that plainly instead of forcing one.Research has diminishing returns, and AI tools make it dangerously easy to keep going past the point where more research changes the decision. The tenth follow-up question rarely moves the recommendation. It just delays acting on the one you already have.
| Stakes | Time-box | What that buys you |
|---|---|---|
| Low, reversible, under ~$100 | 5-10 minutes | One framed question, one source check. Act on it. |
| Medium, some cost or commitment | 20-30 minutes | Framed question, source check, full follow-up ladder, short memo. |
| High, hard to reverse, real money or contract | 45-60 minutes, spread over a day or two | Everything above, plus a second pass the next day with fresh eyes before you commit. |
Pick the time-box before you start, based on what's actually at stake, not on how interesting the topic turns out to be. When the timer runs out, you write the memo with what you have. An imperfect decision made on time almost always beats a perfect one made too late to matter.
This guide is a taste of the research chapter in Your Handy AI Guide to Claude and Claude Code. The Mega Pack walks through the same workflow screen by screen, with the source check built in.
One more thing
New guides drop every couple weeks. No spam, just practical AI from someone actually building it.
Three ways to go deeper. Scan the QR or hit the link. Both work.
Free community of small business owners and operators actually using AI.
Custom AI builds and automations for your business. Direct line, no funnel.
The newsletter. Practical AI in your inbox every couple weeks. No spam.