What AI stock research can verify and what it cannot
AI stock research is useful in a narrow, checkable way: it can gather, date, and organize evidence about public companies faster than you can by hand. It becomes harmful the moment its output cannot be traced back to a source.
That is the whole test. Before trusting any AI research result, ask where each number came from, when it was recorded, and whether the company is actually covered by the system answering you. A fluent paragraph with no dates and no sources is not research. It is prose.
What AI adds to stock research
Three things, in practice.
It translates questions. "Profitable AI-exposed large caps" is a sentence, not a filter set. A language model can turn that sentence into concrete rules and run them against structured data, so the research can start in your own words.
It retrieves and assembles. Reading one 10-K takes an afternoon. Pulling the same five figures from forty filings, each with its filing date attached, is exactly the kind of work a machine should do.
It compares on request. Two companies lined up on the same metrics, same definitions, same as-of dates. That view is tedious to build manually and easy to check once built.
What it can verify
A well-built system can verify anything that lives in a dated record. Filed financials from SEC EDGAR. Reported revenue, margins, and cash flow, each tied to a specific filing. Daily closing prices with their dates. Deterministic screen results, where a company appears because its stored data met a stated rule.
With stocks-llm, every figure shown carries a source and an as-of date, and the model is constrained to a catalog of covered companies. When a value is unavailable it is shown as `n/a` rather than estimated. When a company is not covered, the answer says so instead of describing it from memory. Prices are delayed daily-close values, not real-time quotes. The about page documents how the data works.
What it cannot verify
No AI system can verify a forecast, because a forecast is not a fact. It cannot tell you whether a stock is cheap in any final sense, whether a theme will last, or what a company will earn next year. It cannot know your time horizon or risk tolerance, so it cannot give advice worth taking.
There is also a quieter failure mode. A general-purpose chatbot with no data layer will answer stock questions anyway, from training data that may be years old, and it will rarely warn you. Numbers without dates are the tell. If you cannot find the as-of date, assume the number is stale.
How to pressure-test an AI research answer
A few habits catch most problems.
Ask for the source and the date of any figure that matters, and open the primary source when the amount is consequential. For US issuers that usually means the linked SEC filing.
Check coverage before reading absence as meaning. A company missing from a result may simply be outside the catalog, not a company that failed the screen.
Rerun the question phrased differently. A real, data-backed answer is stable under rephrasing. An improvised one drifts.
Treat `n/a` as missing, never as zero.
Where to start
A reasonable first query is one you can inspect: companies with high revenue growth and strong free-cash-flow margins. The result is a dated company list drawn from stored, sourced data. The research begins after the list appears: read the filings, compare the businesses, and drop names the evidence does not support.
stocks-llm is for informational research only, not financial advice. Verify material information independently against primary sources before making an investment decision.