Finance · Research

How to analyze a 10-K with ChatGPT, and trust the numbers

Use ChatGPT on a 10-K without trusting it blindly: where each number lives, how to avoid period and unit errors, and how to trace every figure to its page.

Key takeaways

  • A 10-K reports the same metric in several places, with different scope. Name the section, period and unit in your question.
  • Ask for each figure with the table row or sentence it comes from, and the PDF page.
  • Recompute growth rates and ratios yourself from the quoted inputs.
  • If a metric is not disclosed in the filing, the right answer is "not in the 10-K", not an estimate.

An annual report on Form 10-K is long, dense and consistent in structure, which makes it a good fit for an AI assistant. It also means the same number can appear in four places, each with a slightly different meaning. Here is how to get useful answers from ChatGPT and confirm every figure before it goes into your notes or model.

Where the numbers live in a 10-K

Knowing the structure lets you ask precise questions, and lets you judge whether an answer came from the right place.

You wantLook in
Revenue, net income, earnings per shareItem 8, the consolidated statements of operations
Segment or regional resultsItem 7 (MD&A) and the segment note in Item 8
Why results changedItem 7, Management's Discussion and Analysis
Debt, maturities and covenantsThe debt note in Item 8, and liquidity in Item 7
Share repurchasesItem 5, and the equity note
EmployeesItem 1, under human capital
RisksItem 1A, Risk Factors

Questions that work, and questions that don't

Vague questions invite blended answers. Precise ones get a figure you can check.

  • Too vague"How did the company do this year?"
  • Better"What were total net sales in fiscal 2025 and fiscal 2024, from the consolidated statement of operations? Quote the row and give the PDF page."
  • Too vague"What's their debt?"
  • Better"List the principal amount of each senior note outstanding at fiscal year end, from the debt note, with its maturity year."

The checks that catch most errors

  • Fiscal period. Many companies do not use a calendar year. Apple's fiscal 2025, for example, ended on September 27, 2025.
  • Units. Statement tables are usually "in millions, except per-share amounts". A missing "million" turns a correct answer into a wrong one.
  • Scope. Consolidated or segment? Continuing operations or total? Read the table title.
  • GAAP or not. Adjusted figures are often in the earnings release rather than the 10-K. If the answer quotes an adjusted number, find where it came from.
  • Page numbers. The printed page number at the bottom of a 10-K page is rarely the same as its position in the PDF.

Recompute growth yourself

Models often get growth rates right, and occasionally do not. Recalculating takes a moment once you have both inputs. Apple's fiscal 2025 10-K reports total net sales of $416,161 million, against $391,035 million in fiscal 2024:

(416,161 − 391,035) ÷ 391,035 = 6.4%

Both inputs appear in the same table, so one highlighted row confirms the whole calculation.

Comparing years and companies

Two common requests are this year against last year, and one company against a peer. Both go wrong in the same way: the inputs come from tables with different scope or definitions.

  • Year on year: use the prior-year column in the current filing, not last year's 10-K. Companies restate prior periods for accounting changes and discontinued operations, and the current filing shows the comparable figure.
  • Company against company: check that both use the same fiscal year end, the same unit, and comparable definitions. "Revenue" and "net sales" are usually equivalent; "adjusted EBITDA" rarely is.
  • Segments: companies define segments differently and change them over time. Compare segments only after reading each segment note.

Running the analysis with CiteLLM in ChatGPT

CiteLLM for ChatGPT opens each answer at its source. Attach the 10-K, ask your questions, and every figure comes back with its table row or sentence highlighted on the original page. Calculated figures show the formula with each input highlighted.

You can attach up to 20 PDFs in one review, so comparing two companies, or this year's filing with last year's, keeps each document in its own tab. No filing at hand? Try it with Apple's fiscal 2025 10-K.

Frequently asked questions

Can ChatGPT read a whole 10-K?

It can work with long filings, but it may not see every page at once. Ask about specific sections and periods, and check each answer against the page it cites.

Where do I get a 10-K as a PDF?

SEC EDGAR publishes filings in HTML. Many companies post PDF versions on their investor relations site, and you can print the EDGAR version to PDF.

Is ChatGPT or Claude better for 10-K analysis?

Both handle filings well. What matters more is checking the answers, and CiteLLM works the same way in ChatGPT and Claude.