QWE AI Academy The question every new investor eventually asks: “How do I actually know if a company is worth buying, or if I’m just guessing?” Not the theory – the actual process. Which document you open first, what you type into ChatGPT, and how you know the answer isn’t made up.
This tutorial walks through how to analyze a company before buying stock using AI paired with primary sources. One honest premise up front: ChatGPT will give you wrong information sometimes, and you need a workflow that catches it before you spend money.
The scenario: a stock you’ve never heard of
A friend mentions a mid-cap semiconductor supplier at dinner. You’ve never heard of it. Thirty minutes before bed – is it worth deeper research tomorrow, or not?
Opening ChatGPT and typing “is [company] a good buy?” is the worst possible move. Not because ChatGPT can’t help – it can – but because the framing produces a confident-sounding paragraph you have no way to verify. The better move is a two-track process: use the AI for structure and speed, use official filings for numbers. That tension is what this whole workflow is built on.
Everything below assumes a U.S.-listed public company. Foreign company trading only on a foreign exchange? The entire SEC pipeline below doesn’t apply – those companies don’t file 10-Ks or 20-Fs with the SEC, and you’ll need their home-country annual report instead.
Why raw ChatGPT alone will burn you
The most important fact in this tutorial isn’t a ratio. It’s a hallucination rate.
Roughly 1-in-5 facts – that’s what a 2025 peer-reviewed study found when it tested chatbots on finance questions. Specifically: ChatGPT-4o produced incorrect citations 20.0% of the time, and o1-preview 21.3% (Erdem et al., International Journal of Data Science and Analytics, 2025). Gemini Advanced hit 76.7% in the same test. The model doesn’t flag which answer is wrong. You get a clean sentence and confidence either way.
Before every session: ask ChatGPT which model is answering and whether web browsing is on. On the free tier, the default has historically been GPT-4o with an October 2023 training cutoff – meaning any “recent” financials it produces from memory are at least two years stale. This may have changed since; verify in your settings before relying on it.
The consequence isn’t abstract. Ask “what was Company X’s revenue last quarter?” and the model invents a plausible number. No error message. You get a clean sentence and confidence. That’s why the workflow below always ends with the primary source – not because of paranoia, but because the math on the error rate is real.
Two tabs. That’s the whole setup.
Tab 1: ChatGPT with web browsing enabled if your plan allows. Synthesizer. Question generator. Do not treat it as a source.
Tab 2: SEC EDGAR – the official filings database, free, public. Type the ticker. Pull up the filings history.
On EDGAR, form types are listed by date. Four matter for a beginner:
- 10-K – annual report. The heaviest document. Start with Item 1 (Business Overview) and Item 1A (Risk Factors).
- 10-Q – quarterly. Shorter, more current when the 10-K is stale.
- 8-K – filed when something material happens: CEO leaves, acquisition, big customer loss. Often the first signal something changed.
- DEF 14A – proxy statement. Executive compensation is in here. Most beginners never open it.
These filings carry strict legal liability for false statements. Marketing decks don’t. Press releases don’t. The 10-K does. That legal exposure is what makes them your ground truth – not tradition, not habit.
The catch: large accelerated filers (public float above $700M) have up to 60 days after fiscal year-end to file the 10-K; mid-size accelerated filers get 75 days. Researching in February for a December 31st fiscal year? The latest 10-K may not exist yet. Grab the most recent 10-Q instead.
Four prompts worth using
Every stock-analysis tutorial recycles the same prompts. These four are built differently – each one is designed so you can cross-check the output against the EDGAR tab.
Prompt 1 - Business model in plain English:
"Explain how [Company] makes money. What are its top 3 revenue segments?
Cite the exact section of their latest 10-K. If you don't know, say 'do not know.'"
Prompt 2 - Risk extraction:
"List the top 5 risks disclosed in [Company]'s latest 10-K Item 1A.
Quote directly. Do not paraphrase or invent."
Prompt 3 - Red flag pattern check:
"Compare [Company]'s revenue and free cash flow over the last 3 fiscal years.
Flag any year where cash flow diverges from reported earnings."
Prompt 4 - What the docs DON'T answer:
"What questions about [Company]'s business does the 10-K leave unanswered
or address only vaguely? Be specific about which sections are thin."Prompt 4 is the one most beginners never think to ask. It flips the AI from “summarize what’s there” to “find what’s missing” – often the more useful question for an actual investment decision. Turns out the gaps in a 10-K are sometimes more revealing than what’s in it.
The “do not know” instruction in Prompt 1 matters. Without it, models tend to fill silence with plausible-sounding guesses. Forcing an explicit admission of ignorance is a simple way to reduce the confident-sounding fabrication that makes the 20% hallucination rate so dangerous in practice.
The verification pass
Four ChatGPT outputs. Don’t act on any of them yet.
Open the 10-K in Tab 2. For every specific number ChatGPT quoted – revenue figures, segment percentages, risk factors – Ctrl+F search the actual filing. ChatGPT said “data center segment was 42% of revenue”? Find the exact table. Number doesn’t appear, or appears differently? That output is contaminated. Mark it and move on.
How often does this happen? Roughly consistent with that 20% figure from the finance study – so across ten specific claims, expect at least one or two that don’t survive the check. Anyone who tells you “just use ChatGPT for stock research” without a verification step is describing a workflow where you eat that error rate silently.
An honest question worth sitting with
Here’s something worth pausing on: if this workflow requires you to open the 10-K anyway and Ctrl+F through it, how much is the AI actually saving you? Probably more than you’d expect – the AI’s real value here isn’t replacing the filing, it’s telling you where to look and what questions to bring to it. The 10-K for a mid-cap company can run 150+ pages. Prompt 4 alone can cut your reading time significantly by surfacing the thin sections first.
What this workflow cannot do
Even done perfectly, this gives you a factual foundation – not a buy signal.
It confirms the company is real, its finances are what management claims, and its stated risks are documented. What it doesn’t give you:
- Whether the current stock price is fair – that requires valuation work (DCF, comparables) beyond this tutorial’s scope
- Whether the industry is about to shift – LLMs are weak at forward-looking industry dynamics, full stop
- Whether management is trustworthy – proxy filings help at the margins, but reputation research lives outside EDGAR
There’s a temptation to keep asking the AI questions until it produces a confident recommendation. The primary-source discipline is the point – not the AI’s opinion, which has no legal exposure and no memory of being wrong.
FAQ
Can I skip SEC EDGAR and just use Yahoo Finance?
You can. But Yahoo aggregates from the same filings and occasionally lags or misreports. EDGAR is the source. For a beginner spending 30 minutes on one company, opening EDGAR once takes 90 seconds and removes a whole class of error. Not a lot of downside.
Which ChatGPT model should I use for this?
Any model with web browsing enabled beats any model without it – browsing lets it pull current filings rather than answer from stale training data. If you’re on the free tier, the default has historically been GPT-4o with an October 2023 training cutoff (as of early 2025; this may have changed). Anything post-2023 that comes from memory – not from a live search – is unreliable for current financials. Turn browsing on before doing this seriously, or the whole verification workflow is fighting an uphill battle from the start.
Does this work for foreign stocks or crypto?
Foreign companies on U.S. exchanges file a 20-F on EDGAR – same idea, different form name. Foreign companies that trade only on foreign markets? EDGAR won’t have them. You’ll need the annual report from their home-country regulator, which varies by country, language, and disclosure standards. Crypto tokens have no equivalent primary-source filing at all – no 10-K, no regulatory equivalent, no legal liability for accuracy – which is one structural reason the same workflow doesn’t transfer. That’s not a knock on crypto; it’s just a different category of asset with different due-diligence tools.
Your next step
Pick one U.S.-listed company you’ve been curious about. Set a 20-minute timer. Run the four prompts above, then verify each output against the 10-K on EDGAR. Count how many specific claims survived. That number – your personal verification baseline – tells you more about how to use this tool than any tutorial can.