Research Anything with AI (and Trust the Answer)

Turn one real question into a one-page brief where every claim has a link you've checked.

Level 2 · Assistant Free · 30 minutes · Updated Oct 30, 2026

Asking AI to “look something up” is easy. Getting an answer you’d stake your name on is not. In the next 30 minutes you’ll turn one real question into a one-page brief where every claim has a link you’ve checked, and then set it up so a fresh update is waiting for you every Monday.

Remember the intern: smart, fast, first day on the job. They’ll happily come back with twelve links in ten minutes. Your job today is learning how to send them out properly, and how to check what they bring back.

What you’ll get


The result	A one-page research brief (findings + sources + what’s still unclear), and a weekly update that runs on its own
Level	Level 2 (an assistant that searches the web). The weekly update edges into Level 3
Loop step you’ll practice	“You check.” How to tell whether a source holds up
Time	30 minutes for the brief, 15 minutes to set up tracking, then 5 minutes a week

What you need

Claude Pro (or above). We’ll use two features: Research and scheduled tasks. Claude searches the web on its own when it needs to; free accounts can search but don’t have Research or scheduled tasks. Other tools work too; see the end.
A free account on one other AI tool (ChatGPT, Gemini or Perplexity) for a quick cross-check in Step 5.
A real question. We’ll use this one throughout: What are the most notable recent success stories of one-person companies? Swap in your own when you’re ready.

The three rules

Everything below comes down to three rules. Each one is an action, not a slogan.

Rule	What you actually do	Why
Reliable	Ask for primary sources first (official announcements, filings, original data, the founder’s own numbers) and make the AI label every source by type and date	Every retelling bends the facts a little, and old news gets passed off as new
Complete	Have the AI split the question into sub-questions before it searches, then ask what it might have missed. Cross-check the key points in a second tool	AI tends to find the loudest answer, not the whole picture
Sourced	Every claim gets a link. You open the three most important ones yourself. Every number gets traced to where it first appeared	A link existing is not the same as the link saying that

One more thing to keep in mind: when AI says “I couldn’t find anything,” that doesn’t mean nothing exists. It can’t see paywalled reports, your company’s files, or sites that block it.

Part 1: One-time research (30 minutes)

Six steps, from a vague question to a brief you can forward without worrying. Copy the prompts as they are, then swap in your own topic.

Step 1. Write a proper brief (3 minutes)

Most bad research starts with a one-line question. “Find me successful one-person companies” leaves the intern guessing: successful by what measure? Since when? Does a founder with two contractors count?

Open a new chat in Claude and paste this, filled in:

I need to research a question. Don't search yet, just read this.

Question: What are the most notable recent success stories of one-person companies, and what do they have in common?
Why I need it: I'm writing a short briefing for people thinking about starting a solo business.
Time range: the last 12 months.
Scope: worldwide; English-language sources are fine.
Definitions:
- "One-person company" = one founder, no full-time employees (freelancers and contractors are OK).
- "Success" = at least $1M in annual revenue, or an acquisition.
What I already know: solo founders often share revenue numbers on social media, and I'm not sure how far to trust them.

The Definitions line does most of the work here. Without it, you’ll get a list of famous startups that had 30 employees.

The five parts of a good research prompt

When you switch to your own topic, check your brief against these five parts. Missing one is the most common reason research comes back vague.

Part	What it answers	Example from above
Question	What exactly do you want to know?	Notable recent success stories of one-person companies, and what they share
Purpose	What will you do with the answer?	A short briefing for people thinking about going solo
Scope	Which time period, which places, which languages?	Last 12 months, worldwide, English sources
Definitions	Which words could be read two ways?	“One-person” = no full-time employees; “success” = $1M+ revenue or an acquisition
Output	What should the answer look like?	A table with links, source types and dates, then the patterns

You don’t need magic phrases like “You are a world-class analyst.” A clear brief beats a clever opening line.

Step 2. Split the question before searching (4 minutes)

Screen recording: asking for the plan, then editing it (coming soon)

Don’t let the intern run off yet. Ask for a plan first:

Before you search, break this question into 4–6 sub-questions.
For each one, list the types of sources you'd use.
Don't search yet. I'll review the plan first.

Read what comes back. Cross out what you don’t need and add what’s missing. For this topic, a good plan should include a sub-question like “Which of these revenue numbers can be confirmed by someone other than the founder?” If it’s not there, add it. This is the step that makes your research complete.

Step 3. Run the research with strict rules (10 minutes)

Screen recording: choosing Research from the + menu, sending the prompt (coming soon)

In Claude, there’s no web search switch anymore: Claude decides on its own when to search. To start a full Research run, click + next to the message box and choose Research, or start your message with /deep-research. Then send:

Go ahead with the plan as I edited it. Rules:
- Prefer primary sources: company or founder announcements, acquisition notices, interviews in established outlets, public revenue dashboards.
- For every claim, give: claim | link | source type (primary / established media / blog / self-reported) | date published | confidence (high / medium / low).
- If a revenue number only comes from the founder, label it "self-reported".
- If you can't find something, say so. Don't fill gaps with guesses.
Output: a table of 8–12 companies, then 3–5 patterns they share.

Research usually takes 5–15 minutes and comes back as a long report with links throughout. Get a coffee.

Just ask, or use Research?

Think of it as asking the intern a quick question in the hallway versus sending them off to write a report.

	Just ask (a normal message)	Research
Good for	One or two specific facts: “When was X acquired?” “When does this rule take effect?”	Broad questions with many angles, like the one in this guide
How many sources	One or two pages would answer it	You’d need to open lots of pages and compare many things
Time	Seconds to a minute	Usually 5–15 minutes, longer for complex topics
Usage	Low	Roughly 5–10 times a normal search

Rule of thumb: if you could answer it yourself with one or two web pages, just ask. If you’d need to open a dozen, use Research. Not sure? Just ask first, and switch to Research if the answer is thin or one-sided. The follow-ups in Steps 4 and 5 are always plain messages. And for specific facts, plain search is often more accurate: in one recent side-by-side test, a quick standard search got the core facts right where several research modes slipped.

Step 4. Ask the two uncomfortable questions (3 minutes)

Two follow-ups:
1. What evidence goes against these patterns? Include solo companies that failed or shut down, if you found any.
2. Which claims rely on only one source? List them.

The first question fights survivorship bias, which is a real problem with a topic like “success stories.” The second one gives you your checking list for Step 6.

Step 5. Cross-check in a second tool (4 minutes)

Pick the two claims that matter most to your brief. Ask a different tool (ChatGPT, Gemini or Perplexity, free versions are fine):

Is it true that [company] [claim, e.g. was acquired in 2026 / makes $X a year]? Give me your sources.

Same answer, different sources? Good. Different answers? Mark that claim as uncertain. Don’t pick the one you like better.

Step 6. Check the links, then write the brief (6 minutes)

Screen recording: opening three sources and checking them against the claims (coming soon)

Open the three most important links yourself. For each one, ask:

Does the page open?
Does it actually say what the AI says it says?
For numbers: is this the original source, or someone repeating it?

Delete anything that fails. Then ask Claude to write it up:

Turn this into a one-page brief, using only the claims I've kept: [paste the rows you kept]
1. Bottom line, in 3 sentences
2. Table: company | what they do | key number | source | self-reported?
3. Patterns
4. What's still uncertain or missing

That last section is not optional. A brief that says what it doesn’t know is one people can trust.

Part 2: Keep tracking it (15 minutes to set up, 5 minutes a week)

Topics like this move every week. Instead of redoing Part 1 every month, hand the intern a standing assignment: check every Monday and report only what’s new.

Step 1. Start from a run that worked

Only schedule something you’ve already done well once by hand. Schedule a sloppy prompt and you’ll just get sloppy results every week.

Step 2. Rewrite the prompt for tracking

Scheduled runs start fresh each time and can’t ask you questions, so everything has to be in the prompt:

Every Monday at 8:00, search for news from the past 7 days about one-person companies (one founder, no full-time employees) that announced revenue milestones, funding or an acquisition.
- Only report what's new in those 7 days.
- For each item: claim | link | source type (primary / established media / blog / self-reported) | date | confidence.
- Label founder-only revenue numbers as "self-reported".
- Maximum 5 items, most important first.
- If nothing significant happened, reply only: "No significant updates this week."

The last line matters more than it looks. Without it, the intern will scrape together something every week just to have something to say.

Step 3. Schedule it in Claude

On Pro and above, paste the prompt into a new conversation and say you want it to run weekly. Claude will suggest a name, a schedule and the instructions; check them and click Schedule. You can pause, edit or run it right away from the Scheduled page. New scheduled tasks run in the cloud, so your laptop can be closed. Each run counts toward your usage limit.

Step 4. Spend 5 minutes on each update

Open the one or two most important links and check them, just like Step 6 above. Fresh news is where mistakes are most common, and also what people are most likely to forward without reading.

Step 5. Tune it once a month

Are the sources getting one-sided? Are you getting too much noise, or missing things? Adjust the prompt. Once it’s stable, you can save the whole method as a Skill, so the next topic you track starts from your proven instructions instead of a blank page.

Why you can’t skip the checking

This isn’t paranoia. Researchers have measured it:

A 2025 study checked AI search answers sentence by sentence and found that roughly 25% to 47% of statements weren’t supported by the sources the tool itself cited (DeepTRACE, Salesforce AI Research).
A 2026 study found that 3% to 13% of links given by AI research tools probably never existed, and that deep research tools invent links more often than plain search does (arXiv 2604.03173).

More links and longer reports don’t automatically mean more reliable. Opening three links takes five minutes, and it’s the difference between a brief people trust and one they quietly stop reading.

Checklist

Before you send your brief to anyone:

Every claim has a link, and the link opens
I read the original for the three most important claims, and it really says that
Every number is traced to where it first appeared, not to someone repeating it
At least some sources are primary, not all blogs and vendor posts
Every source has a date; nothing old is presented as new
Self-reported numbers are labeled as self-reported
The brief says what’s still uncertain or missing

What you just learned

On the surface you looked something up. Underneath, you practiced the two steps of the loop that most people skip: giving context (Steps 1–2) and checking (Steps 4–6). The AI did the searching; you did the judging.

The same method works for: competitor research, tracking a regulation, preparing for a client meeting, vetting a supplier. Change the brief in Step 1, and keep everything else.

Using a different tool?

The prompts work anywhere. Only the buttons change.

Tool	For Part 1 (Step 3)	For Part 2 (scheduling)
Claude	+ → Research, or type /deep-research (Pro and above)	Scheduled tasks (Pro and above)
ChatGPT	Deep research	Scheduled Tasks (paid plans)
Gemini	Deep Research	Scheduled Actions (paid plans)
Perplexity	Research mode	Not covered here
Microsoft 365 Copilot only	Web search in Copilot Chat; follow Steps 1–6 as written	Set a Google Alert on your keywords, and run the tracking prompt by hand every Monday

Features and plan names change often. This table was checked in October 2026.