Research Like an Analyst

Use ChatGPT's three research levels to build a verified Decision Brief.

Introduction

30 Second Summary

Every day, people ask AI for advice on career moves, investments, and big decisions. The answers sound confident and thorough, but confidence is not the same as accuracy.

In this project, you will investigate a real career question using three levels of ChatGPT research. You will produce a professional Decision Brief that traces a recommendation back to real evidence and honestly states what you do not know.

What You'll Build

You will have a Decision Brief you could hand to a mentor, showing exactly what you recommend, the evidence behind it, the evidence against it, and the specific claims you personally verified by clicking through to the original sources.

By the end of this project, you'll have:

A completed Decision Brief with a clear recommendation, the evidence that supports it, conflicting findings, and an honest confidence level.
Firsthand experience with three research levels (built-in knowledge, search, and deep research) and the judgment to know when each one is the right tool for a question.
A tracking table of AI-generated claims you personally verified by clicking through to original sources, proving you can check rather than just trust.
Secret Mission: Stress-test your brief by finding the strongest counterargument to your recommendation and deciding whether it changes your conclusion.

Are there any prerequisites?

You need a ChatGPT account and should be comfortable using it for basic conversation. The Free plan works for this project (5 deep research queries per month), though ChatGPT Plus at $20/month is recommended for 25 queries per month.

Before We Start

Before you open any tools, take a moment to lock in what you are about to build and why it matters.

In this project, you will investigate a real career question using three levels of ChatGPT research. Each level gives you a different quality of answer to the same question. By the end, you will have a professional Decision Brief that traces a recommendation back to real evidence.

Set Up Your Research Workspace

In this project, you will ask ChatGPT the same career question three different ways. Each way taps a different research level, from instant recall to live web search to a full multi-source investigation.

Each level needs its own clean conversation so prior messages do not influence the results. In this step, you will sign in to ChatGPT and confirm your account is ready for all three levels.

In this step, get ready to:

Sign in to your ChatGPT account at chatgpt.com.
Confirm your plan includes access to deep research.
Plan your three-chat workspace for side-by-side comparison.

Sign in to ChatGPT

Open your web browser.
Navigate to chatgpt.com.
Sign in with your ChatGPT account credentials.

You should see the ChatGPT home screen with a message composer at the bottom of the screen.

Can't sign in?

Make sure you are using the email address you registered with. If you have forgotten your password, use the password reset link on the sign-in page.

If you do not have an account yet, create one for free at chatgpt.com.

Help me sign in to ChatGPT.

Confirm access to deep research

ChatGPT offers a feature called deep research that runs a multi-step investigation across many web sources. You will use it in Step 4, so let's confirm it is available in your account now.

Start a new chat.
In the message composer at the bottom of the screen, click the + icon to open the tools menu.
Look for Deep research in the list of available tools.

✔️ I can see Deep research

Your account has access to deep research. You are ready to move on.

Your deep research limits

On the Free plan, you get 5 deep research queries per month. On Plus, you get 25. This project uses exactly one deep research query, so either plan works.

ⓧ I don't see Deep research

Deep research availability depends on your plan, region, and rollout status. Try these steps before moving on.

Type /Deepresearch directly in the message composer to see if it activates the feature.
Confirm you are signed in to the correct account.
If deep research is still not available, you can follow along with every other step in this project. When you reach Step 4, the content will describe the expected output so you can continue building your Decision Brief.

Help me find deep research in ChatGPT.

Keep three chats open

You will use three separate chats in this project, one for each research level. Keep all three open in separate browser tabs so you can compare results side by side at the end.

Why three separate chats?

ChatGPT uses earlier messages in a conversation to shape its next response. If you ask the same question three times in one chat, each answer builds on the last. Starting fresh each time removes that influence, so you can see what each research level produces on its own.

You should now have chatgpt.com open with a new, empty chat ready. If you are on Plus, you confirmed that Deep research appears in your tools menu.

Your workspace is ready. In the next step, you will ask ChatGPT a career question using only its built-in knowledge and see how confident its answer sounds.

Get the Instant Answer (Built-in Knowledge)

Your research workspace is set up and your new chat is ready. Now it's time to ask the career question that drives this entire project.

When you ask ChatGPT a question in a normal chat, it draws on built-in knowledge from its training data, which can be months or years old. You need to capture this baseline answer first so you can judge later whether current information changes the picture.

In this step, get ready to:

Ask ChatGPT your career question and get a baseline answer without search.
Challenge ChatGPT to identify which of its own claims are most vulnerable to change.
Create a tracking table to record specific claims for verification in later steps.

Ask for the baseline recommendation

In a normal chat, ChatGPT answers from what it learned during training. No web search, no external tools, just what it already "knows."

In your open chat at chatgpt.com, copy and paste this prompt into the message composer:

Should I learn prompt engineering or traditional software development to maximize my career value over the next two years? Give me a direct recommendation with your reasoning.

Before sending, make sure you have not selected any tools from the tools menu (the + icon in the composer area).
Press Enter to send the prompt.

Why send this without search?

You want to isolate what ChatGPT "knows" from what it can find online right now. By keeping search and other tools disabled, you get a pure baseline.

Any gaps or outdated claims in this answer will become visible when you compare it against search results in the next step.

Read the response carefully. As you read, pay attention to these questions:

How confident does the answer sound?
Does it cite any sources or link to any data?
Does it mention any specific dates, salary figures, or job market data?

Copy or save this response somewhere you can reference later, such as a separate document, a note, or a text file on your computer.

Seeing a globe icon next to the response?

A globe icon means ChatGPT automatically used search. This happens when ChatGPT decides the question would benefit from web information.

Start a new chat and try again. Rephrase slightly or preface your prompt with "Without searching the web:" to discourage automatic search.

Still having trouble?

Expose the assumptions behind the answer

The baseline answer likely sounded confident. But some of those claims depend on data that changes every few months, like hiring trends and salary ranges.

Before you send the next prompt, look at ChatGPT's answer and pick one or two claims you think are most likely to be time-sensitive.

In the same chat, copy and paste this follow-up prompt and press Enter:

What specific claims in your answer would change if the job market shifted in the last 6 months? List them.

What does this follow-up do?

This forces ChatGPT to audit its own answer. It identifies which claims are built on assumptions about the current job market rather than stable facts.

The claims it flags as potentially outdated are exactly the ones worth verifying with real, current data.

Read ChatGPT's response and note which claims it flags as potentially outdated.
Compare the flagged claims to the ones you predicted. Did ChatGPT surface anything you did not expect?

Create your claim tracking table

You now have a baseline answer full of specific claims. To track whether those claims hold up against current data, you need a simple system.

Open a notebook, document, or a new note on your computer.
Create a table with three columns: Claim, Source, and Status.

Here is an example of the format:

Claim                          | Source | Status
-------------------------------|--------|---------------------------
Example: "Prompt engineering    |        |
salaries average $120k"        |        |
                               |        |
[your claim from ChatGPT]      |        |
[your claim from ChatGPT]      |        |
[your claim from ChatGPT]      |        |

What goes in each column?

Claim: a specific, checkable statement from ChatGPT's baseline answer. Salary ranges, job demand statements, and growth predictions work best.
Source: leave this blank for now. You will fill it in when you search for current evidence in the next step.
Status: will be one of Confirmed, Contradicted, or Uncertain. Leave this blank for now too.

Go back to ChatGPT's baseline answer and pick 3-4 specific claims.
Write each claim in the Claim column of your tracking table.
Leave the Source and Status columns blank.

Pick claims you can actually check

The best claims to track are ones with numbers or timeframes, like salary ranges, percentage growth figures, or statements about which skills employers are hiring for right now.

Vague claims like "software development is valuable" are hard to verify. Specific claims like "prompt engineering roles grew 300% in 2025" can be checked against real data.

Look at your tracking table. You should have 3-4 specific claims written down with the Source and Status columns blank, ready for the next step.

You now have a confident answer from ChatGPT, a list of its own admitted assumptions, and a tracking table loaded with claims to test. Next up, you will ask the same question with search enabled and see how the answer shifts when current data enters the picture.

Search for Current Evidence

You now have a baseline answer and a tracking table full of claims that ChatGPT itself flagged as potentially outdated. The natural next question is whether current data actually supports them.

ChatGPT's search feature pulls live information from the web and returns source links alongside its answer. In this step, you will ask the same career question with search enabled and find out whether the confident claims from Step 2 hold up against current data.

In this step, get ready to:

Ask the same career question with search enabled and compare the new response to your baseline.
Evaluate the sources ChatGPT cites by type, date, and data quality.
Verify a specific claim by clicking through to the original source.

Ask the same question with search enabled

Start a new chat in ChatGPT. Keep your first chat (the baseline from Step 2) open in another tab so you can compare later.
Copy and paste this prompt into the new chat:

Search the web for current information: Should I learn prompt engineering or traditional software development to maximize career value over the next two years? Include recent job market data, salary trends, and employer demand. Cite your sources.

What makes this prompt different?

Starting your message with "Search the web" explicitly tells ChatGPT to use its search feature instead of relying on training data alone.
Asking for "recent job market data, salary trends, and employer demand" directs the search toward concrete, verifiable figures rather than general advice.
"Cite your sources" ensures the response includes links you can click through and check.

Send the prompt and wait for the response to finish generating.
Look for the globe icon next to the response. This confirms ChatGPT used its search feature to pull live web data.

You should also see inline citations throughout the response. Each one links to a real web source you can visit.

Don't see the globe icon?

ChatGPT sometimes decides a question does not need live search results. To force it, click the + icon in the composer to open the tools menu, then select the Search icon and re-send your question.

You can also type / in the prompt and select Search from the pop-up menu.

Why isn't ChatGPT using search when I ask it to?

Evaluate the sources ChatGPT cites

Each citation link in the search-backed response points to a real web page. Not all sources carry equal weight. A government labor statistics report and a personal blog post both show up as links, but they tell you very different things.

Click 2-3 of the citation links in the response. For each source, note the following:

Return to your tracking table from Step 2. For each claim, compare what the search-backed response says to the original baseline answer.
Update the Status column for each claim: mark it as Confirmed (search agrees with baseline), Contradicted (search disagrees), or Uncertain (search does not address it clearly).
Note any new claims that appeared only in the search-backed response. Add them to your tracking table as new rows.

Why does source type matter?

A salary figure from a national labor statistics bureau uses large sample sizes and standardized methods. A figure from a recruiter's blog post might cherry-pick data to support a narrative.

Knowing the source type helps you gauge how much weight each claim deserves. Two claims with the same confidence level can mean very different things depending on where the data came from.

Verify a claim against its original source

Seeing a citation link next to a claim feels reassuring. But a link does not guarantee accuracy. ChatGPT sometimes summarizes a source differently from what the source actually says.

Before you click through, pause and predict: do you think the source will say exactly what ChatGPT reported, or will there be differences?

Pick one specific data point from the search-backed response. Choose something concrete: a salary figure, a percentage growth claim, or a hiring trend.
Click through to the original source that ChatGPT cited for that claim.
Read what the source actually says. Ask yourself two questions:

If the source does not support the claim, update your tracking table and mark that claim as Unverified.
If the source confirms the claim, mark it as Confirmed and write down the source name or URL in the Source column.

Your tracking table now has source information filled in for each claim, with at least one source clicked through and checked against what ChatGPT reported.

What did this exercise reveal?

Most people stop at the citation link. If ChatGPT says "according to [Source]" and includes a link, it feels verified. This exercise proves that the extra step of actually reading the source matters.

A citation makes a claim look credible. Only clicking through makes it actually credible.

Source behind a paywall or removed?

If you cannot access the original source, try a different claim from the response and click through to that source instead. Choose a data point from a source that loads fully.

If the page has been removed entirely, mark that claim as Uncertain in your tracking table. A dead link means you cannot confirm or deny the claim.

How can I verify a claim when the original source is behind a paywall or has been taken down?

Your tracking table now tells a clearer story. Some claims held up, some shifted, and at least one source may not have said what you expected. Next, you will run deep research to see what happens when ChatGPT synthesizes across dozens of sources instead of a handful.

Go Deep (Multi-Source Research)

Your search results gave ChatGPT's baseline answer a reality check. Some claims held up against current data, others shifted, and you now have source links you can trace back to real evidence.

Search returns whatever ranks first, not what is most thorough. In this step, you will use ChatGPT's deep research to synthesize across many sources and produce a single structured report with citations you can evaluate.

In this step, get ready to:

Launch a deep research session on the career question.
Examine the structured report, its sources, and its activity history.
Compare findings across all three research levels and update your tracking table.

Launch a deep research session

Your first chat used ChatGPT's built-in knowledge. Your second used search to pull live web results. Now you will use deep research, which goes far beyond a single web lookup.

What is deep research?

Deep research is agentic. It actively plans a multi-step research process, searches the web, evaluates sources, refines its queries, and synthesizes findings into a structured report. It does not simply return a list of links.

The output is a full report with a table of contents, cited sources, and an activity history showing how the research progressed. This process may take 5 to 30 minutes because ChatGPT is reading and cross-referencing across many pages.

Return to ChatGPT and start a new chat. This is your third chat. Keep Chat 1 (baseline) and Chat 2 (search) open so you can compare all three later.
Click the + icon in the composer area to open the tools menu.
Select Deep research.

Don't see deep research?

You can also type /Deepresearch directly in the message composer and select it from the pop-up menu.
Free plan users get 5 deep research queries per month. If you have already used all 5, your queries switch to a lightweight version. This project uses exactly 1 deep research query.
Deep research availability depends on your plan, region, and rollout status. If it does not appear at all, check the OpenAI Help Center for current availability.

Still stuck? Help me figure out why I can't access deep research in ChatGPT.

Type this prompt into the message composer:

I need to decide whether to invest in learning prompt engineering or traditional software development to maximize career value over the next two years. Research this question using recent job market data, salary surveys, employer hiring trends, and industry forecasts. Produce a structured report with: key findings, evidence for each side, conflicting evidence, a recommendation, confidence level, and what remains unknown. Cite all sources.

Send the prompt. ChatGPT will show you a proposed research plan before it begins. It may also ask clarifying questions to confirm your goals.
Review the plan to make sure it covers job market data, salary information, and industry trends.
If the plan looks reasonable, let it proceed.

What is the research plan?

Before deep research starts browsing, it shows you its plan. This is a key difference from search. You can review and modify the plan so the final report stays aligned to what you are trying to accomplish.

Good research starts with a plan, not just a question. This is true for humans and for AI.

Deep research takes time. Your report may take 5 to 30 minutes to complete. You can leave the tab open and come back when it finishes.

Examine your deep research report

When the research completes, the report opens in a fullscreen view. This is a structured document, not a chat reply.

Scroll through the table of contents to see how the report is organized.
Scroll to the bottom and find the sources used section. Count how many sources were consulted.
Look at the activity history to see how the research progressed. Notice the sequence of searches, source reads, and query refinements ChatGPT performed.
As you read the findings, note whether the report flags any conflicting evidence or uncertainty.

What should I pay attention to?

Focus on two things. First, count the sources. Deep research should have consulted far more sources than your single search query did. Second, look for places where the report says evidence is mixed or conflicting. Those are the most honest parts of the report.

Compare results across all three levels

Before you start comparing, take a moment to predict. Based on what you saw from ChatGPT's built-in knowledge and search, do you think deep research will mostly confirm your current findings or reveal significant new information?

Open your tracking table from the previous steps.
For each claim in the table, check whether deep research confirms, contradicts, or adds new nuance compared to what you found in Step 2 and Step 3.
Update the Status column for every claim. Use Confirmed if all three levels agree, Contradicted if they disagree, or Uncertain if the evidence is mixed.
Pay special attention to claims where all three levels disagree. The baseline said one thing, search said another, and deep research found a third perspective. Write a short note next to each disagreement describing what differed.

Why are disagreements the most valuable finding?

When three different research methods produce three different answers, you have found a question where confidence should be lowest. That is not a failure. It is the most useful information your research produced.

A Decision Brief that honestly states where the evidence conflicts is far more trustworthy than one that pretends every source agrees.

Your tracking table should now show a clear picture. Each claim has a status based on all three levels of research, and disagreements are noted with a short explanation.

Not sure whether a claim is confirmed or contradicted?

If the deep research report does not address a specific claim from your table, search for keywords from that claim within the report. If it is still not mentioned, mark the claim as Uncertain and note that deep research did not cover it.

If two sources in the deep research report disagree with each other about the same claim, that is a contradiction worth noting, even if the report does not call it out explicitly.

Help me compare my three research levels and figure out which claims are confirmed or contradicted.

You now have the raw material for a real decision. Three research levels, one tracking table, and honest notes about where the evidence agrees and where it does not. Next up, you will turn this into a professional Decision Brief.

Build Your Decision Brief

You now have raw material from three research levels: a confident baseline, a search-backed answer with sources, and a deep research report with citations across dozens of sources. Raw material is not a decision.

A Decision Brief forces you to weigh conflicting evidence, take a position, and be honest about what you do not know. This is the structure that separates someone who googled a question from someone who researched it.

In this step, get ready to:

Draft a Decision Brief using your tracking table and a structured prompt.
Verify an unverified claim by clicking through to its original source.
Set your own Confidence Level based on the evidence you have gathered.

Draft your Decision Brief with a structured prompt

Your tracking table holds the evidence. Now you need ChatGPT to synthesize it into a structured document that takes a clear position.

Start a new chat in ChatGPT. This is your fourth chat. Keep the other three open so you can compare results.
Copy your full tracking table (all claims, sources, and statuses) and paste it into the message composer.
Below your pasted tracking table, add the following prompt:

Using the research evidence I have gathered (pasted above), create a Decision Brief with these sections: Question, Key Findings (3-5 bullet points), Evidence For the recommended option, Evidence Against the recommended option, Conflicting Evidence and Tradeoffs, Recommendation, Confidence Level (High / Medium / Low with justification), What Remains Unknown, and Sources. The recommendation must trace back to specific evidence. Do not hedge everything. Take a position.

Send the message.

What is a Decision Brief?

A Decision Brief is a structured document that traces a recommendation back to evidence. It is the format professionals use when they need to show their reasoning, not just their conclusion.

The prompt asks ChatGPT to produce nine distinct sections. The most important ones are Conflicting Evidence (where sources disagree), What Remains Unknown (gaps no source filled), and Confidence Level (how much you trust the recommendation). These sections are what separate research from opinion.

Before you read the response, pause. Do you think ChatGPT's recommendation will match the deep research report, or will it shift now that you have given it your tracking table with noted disagreements?

Read the Decision Brief that appears. You should see all nine sections laid out with your evidence organized into a clear structure.

Check the Recommendation section. Does it actually follow from the evidence listed above it, or did ChatGPT pick the safe, hedge-everything answer?
Check the Sources section. Are the sources listed ones you recognize from your three chats, or did ChatGPT introduce new ones you have not seen before?

Brief looks generic or vague?

If the Decision Brief reads like a generic pros-and-cons list without specific data points, your tracking table may not have included enough detail. Copy a few specific data points from your deep research report (salary figures, growth percentages, or hiring trend quotes) and paste them into the same chat. Then ask ChatGPT to revise the brief using the new evidence.

Still getting a hedge-heavy answer with no clear recommendation?

Run your trust test on an unverified claim

Your Decision Brief looks polished. It has sections, citations, and a clear recommendation. But at least one claim in that brief is something you have not personally checked against the original source. That is your trust test.

Read through the Decision Brief and find one specific claim you have NOT yet verified by clicking through to the original source. Look for data points like salary figures, percentage growth claims, or hiring trend statements.
Click through to the cited source for that claim. Read what the source actually says.

Ask yourself three questions. Does the source actually say what ChatGPT attributed to it? Is the data from 2024, 2025, or 2026? Is the source a credible publication or an opinion piece?

If the claim holds up, return to your tracking table and mark that claim as Verified in the Status column.
If the claim does not hold up, update the Decision Brief to reflect what the source actually says. Correct the data point or rephrase the finding.
Either way, add a one-line verification note at the bottom of your brief using this format:

Verification check: [[VERIFIED_CLAIM="the specific claim you checked"]] was [[CLAIM_STATUS="confirmed or corrected"]] against [[CLAIM_SOURCE="the source name or URL"]].

Why does this matter?

A citation does not automatically make a claim trustworthy. Deep research can still misattribute data, misinterpret a source, or pull from a source that has since been updated. The verification note proves you checked. It also trains the habit of spot-checking before trusting polished-looking AI output.

Set your own Confidence Level

ChatGPT assigned a Confidence Level to your brief. That is ChatGPT's assessment. The confidence level is yours to set, not ChatGPT's.

Open the Confidence Level section of your Decision Brief.
Look at your tracking table. If it shows major disagreements across the three research levels, the confidence should be Low regardless of what ChatGPT wrote. If the evidence consistently pointed in one direction across all three levels, raise it to High.
Edit the Confidence Level and its justification to reflect your honest assessment. Write one sentence explaining why you chose that level, referencing specific evidence from your tracking table.

Confidence is about the evidence, not the recommendation

A strong recommendation can have Low confidence if the supporting evidence is thin or contradicted. A weaker recommendation can have High confidence if every source agrees. Confidence measures how much you trust the evidence, not how much you like the answer.

Review the complete brief one final time. Confirm it has all nine sections: Question, Key Findings, Evidence For, Evidence Against, Conflicting Evidence and Tradeoffs, Recommendation, Confidence Level, What Remains Unknown, and Sources.
Confirm the verification note is at the bottom of the brief.
Save your completed Decision Brief. You can copy the text into a document, or select and copy the full response from the ChatGPT chat.

Missing a section in your brief?

If ChatGPT skipped a section (most commonly What Remains Unknown or Conflicting Evidence and Tradeoffs), send a follow-up message in the same chat asking it to add the missing section. Reference a specific disagreement from your tracking table to give it material to work with.

Need help getting the brief into shape?

You now have a completed Decision Brief with a recommendation that traces back to evidence, a verified claim, and a confidence level you set yourself.

Secret mission

Stress-Test Your Recommendation

Your Decision Brief has a clear recommendation. But a real analyst actively tries to break their own conclusion. Find the strongest counterargument to your recommendation and decide whether it still stands.

Clean Up Your Resources

This project ran entirely in your browser using ChatGPT, so there are no ongoing costs or cloud resources to worry about. Decide whether to keep your chats and documents for future reference, or delete them entirely.

Resources you used:

5 ChatGPT chats (Baseline, Search, Deep Research, Decision Brief, Stress Test).
Decision Brief (saved or downloaded document).
Tracking table (local note or document).

Keep everything running

No action needed. Choose this if you want to revisit your research chats or Decision Brief later.

Your five ChatGPT chats remain in the sidebar of chatgpt.com. You can reopen any of them to review your baseline answer, search results, deep research report, or stress test.
Your Decision Brief and tracking table stay wherever you saved them (a local file, a notes app, or a downloaded document).
None of these resources incur any cost by existing.

Pause - I'll come back to this later

There is nothing to pause in this project. Your ChatGPT chats and local documents do not consume resources while idle.

Make sure your Decision Brief is saved somewhere you can find it. If you only have it inside a ChatGPT chat, download it as a local file so it does not depend on that chat history.
Your tracking table is useful as a reference for how claims shifted across the three research levels. Save it alongside your brief if you have not already.

Delete - I don't want to use this again

Remove all project chats and documents to start fresh.

Delete each of the five ChatGPT chats. In the chatgpt.com sidebar, hover over a chat title, click the three-dot menu icon that appears, and select Delete. Repeat for all five chats (Baseline, Search, Deep Research, Decision Brief, Stress Test).
Delete your saved Decision Brief. If you saved it as a local file, move it to the Trash (macOS) or Recycle Bin (Windows) and empty it.
Delete your tracking table. If it is a local file or note, remove it the same way.

Nice Work!

You just produced a professional Decision Brief on a real career question, built from three levels of ChatGPT research that each revealed something the previous level missed.

You've learned how to:

Produce a Decision Brief that traces a recommendation to specific evidence, flags conflicting findings, and honestly states what remains unknown.
Distinguish when ChatGPT's built-in knowledge, search, and deep research are each the right tool for a question, because you saw the same question produce three different answers.
Verify AI-generated claims against their original sources, proving that a confident answer with a citation is not automatically a trustworthy one.
Secret Mission: Stress-test your recommendation by finding the strongest counterargument, verifying it against its source, and deciding whether it changes your conclusion.

You now know how to create, analyze, and research with ChatGPT. Serious work does not fit neatly into one conversation.

Next up, you will turn ChatGPT from a collection of chats into a workspace.

Ready to quiz yourself?