Using a General AI Chatbot for Stock Research, With Sources
Updated ·5 min read·Reviewed by the StockTools.ai Research Team
- ▸A chatbot answering from memory can be out of date or simply wrong; one reading a page you chose can be checked against that page.
- ▸Every prompt here names a page, asks for quotes and dates, and tells the model to say not found rather than fill a gap.
- ▸On 21 October 2026 our AI Answer Ledger had scored 80 calls on where ten large caps would close 30 days later; 46 picked a direction, and 23 of those were right.
- ▸All 46 said higher, so the 50% hit rate equals the always-higher baseline, and the ledger grades nothing against that baseline until 20 calls have said lower.
Point it at a source, not at its memory
Ask a general chatbot about a company and it answers from whatever it absorbed in training, which may be months old, or from a quick web search it does not show you. The joint investor alert from the SEC's Office of Investor Education and Advocacy, NASAA and FINRA warns that AI-generated information can rest on inaccurate, incomplete or outdated data, can repeat false information spread to move a stock, and can be completely made up even when the input was accurate.
The fix is to give the model a document and make it work from that document. Every page linked below is built from primary sources, mostly SEC EDGAR filings and FINRA data, and prints its dates. If your chatbot can open web pages, paste the address; if it cannot, open the page yourself and paste its text. Either way the answer now has something it can be checked against.
Six prompts that start from a filing
Replace TICKER with the symbol you are researching. 1. The filing list: 'Open https://stocktools.ai/stocks/TICKER/filings. List every 8-K filed in the last 90 days with its date and item numbers, in date order. If an item number is not shown, write not shown. Do not describe any filing you did not see on the page.'
2. Offerings: 'Open https://stocktools.ai/tools/offering-checker/TICKER. Give me the last announced offering and the last executed one, each with its form type and filing date, exactly as the page shows them. If the page says nothing was found in its window, repeat the window dates instead of saying never.'
3. Dilution: 'Open https://stocktools.ai/tools/dilution/TICKER. Summarize the share count history, any warrants and the cash runway, and next to each number give the filing or period it comes from. Mark anything the page labels as an estimate.'
4. Short interest: 'Open https://stocktools.ai/tools/short-interest/TICKER. Report the latest shares short and days to cover with the settlement date they are as of, then tell me how many days ago that settlement date was.' The settlement date is the one that matters, and our short interest dates guide explains why.
5. A recent move: 'Open https://stocktools.ai/tools/why-moving?t=TICKER. Tell me what the page found for the latest session: any 8-K, any halt, any headline, with dates. If it says nothing company-specific is on file, say that and stop.'
6. The annual report: 'Open the company's latest 10-K from https://stocktools.ai/stocks/TICKER/filings. Quote the first two paragraphs of Item 1A, Risk Factors, word for word, and give the filing date. Then list any risk factor headings that mention debt or financing.' Our 10-K guide explains why a search for Item 1A lands on the contents page first, a trap models fall into too.
Check the answer against the page
Spot-check three things in every reply. The dates: each should match a date printed on the page. The numbers: pick one figure and find it on the page yourself. The quotes: a quoted sentence should appear on the page exactly, and a quote you cannot find is a made-up quote.
When a reply contains something the page does not, ask where it came from. A model that cites a page it never opened, or adds a figure with no source, has told you how much to trust the rest of that answer. The prompts above ask for not found on purpose: a model that admits a gap is easier to check than one that fills it.
Where a general AI falls short: what our ledger measured
Our AI Answer Ledger asks a model one fixed, impersonal question about each of ten large caps every week, records the answer with that day's close, and scores it 30 days later against what the price did. Fresh calls are never shown; only scored ones are. As read on 21 October 2026, it had scored 80 calls since 3 August 2026. In 34 the model declined to pick a direction. In the other 46 it picked one, and 23 of those 46 were right: a 50% hit rate.
That 50% needs its baseline. All 46 decided calls said higher, so the hit rate is exactly the share of those windows in which the stock closed higher, which the ledger prints as an always-higher baseline of 50% on the same 46 windows. The ledger itself says it measures the market over those windows, not the model, and it compares the model against that baseline only once 20 or more calls have said lower. None has, so on that comparison it abstains, and so do we.
The windows also overlap. Calls come weekly on the same ten stocks with 30-day horizons, so on one stock two windows a week apart share 77% of their trading days, by the ledger's own measurement. The ledger estimates the 46 decided windows carry about as much information as 17 independent ones, with a 95% interval on the hit rate of 29% to 71%, figures it labels as derived and approximate. Its page updates every week, so read the live numbers there rather than these.
What the ledger cannot tell you
It measures one model, one question, ten large caps and one 30-day horizon. It says nothing about how well a chatbot summarizes a filing, finds an 8-K or reads a table, which are the tasks the prompts above use it for, and it cannot say whether a different model or a different question would do better. A hit rate on direction is also a different thing from being useful: a model that always says higher will score well in any stretch when most stocks rise, without knowing anything.
That is why the prompts above never ask a model which way a stock will go or whether to buy it. A question like that has no answer you can check today, and a confident reply to it is the failure the regulators' alert describes. A question about what a filing says has an answer printed on the page.
FAQ
Can I use a general AI chatbot to research a stock?
Yes, as a reader and summarizer of documents you choose. Give it a page built from filings, ask it to quote and date what it reports, and check one figure yourself. The joint SEC, NASAA and FINRA alert warns that AI output can be inaccurate, outdated or made up.
Why do the prompts tell the model to say not found?
Because a model asked a question it cannot answer from the page tends to fill the gap. Telling it to write not found makes a missing fact visible instead of invented.
How accurate are AI calls on where a stock will go?
Our AI Answer Ledger, read on 21 October 2026, had 46 scored calls that picked a direction; 23 were right. All 46 said higher, matching an always-higher baseline of 50%, and the ledger makes no comparison against that baseline until 20 calls have said lower.
Does StockTools show what an AI thinks a stock will do?
No. The ledger records each call the day it is asked and shows it only after it has been scored 30 days later, beside what the price did. A fresh call is never displayed.
What if my chatbot cannot open web pages?
Open the page yourself, copy its text and paste it into the chat with the same instructions. The answer is then tied to text you can see, which is the point.
Put it to work
Related guides
Sources & further reading
More to learn
Educational only — not financial advice. Concepts simplified for clarity; markets are messier than definitions.