Guide
Why does ChatGPT get the numbers in my PDF wrong?
ChatGPT gets PDF numbers wrong because a PDF stores positioned text and images, not a table of values with their meaning: figures inside charts, scans or complex layouts can be skipped, merged or attached to the wrong period, currency or row. The fix is to give AI the numbers as structured text next to their labels and qualifiers — ideally one fidelity-checked, text-based version of the document that every AI reads instead of re-extracting the PDF.
What a PDF actually gives an AI
A PDF is a print format. It tells a viewer where to draw each character, line and image on a page; it does not say “this cell is Q3 revenue in euros, unaudited.” An AI that reads the file has to rebuild that meaning from positions, and every reconstruction is a guess.
OpenAI’s help center says ChatGPT may not reliably extract exact values from image-based tables, scanned files or complex visual layouts, and recommends uploading spreadsheets or text files when exact values matter. It also notes that reading charts and images inside PDFs (visual retrieval) is offered in specific contexts; elsewhere only the text layer is used. A number that exists only in a chart or an image can simply be invisible.
The typical errors, by document type
| Document | What goes wrong | Why |
|---|---|---|
| Annual or quarterly report | Prior-year figure quoted as current year | Multi-column tables lose their column headers during extraction |
| Bank or account statement | Debits and credits swapped, balances misaligned | Amounts are positioned in columns without explicit labels |
| Price list or spec sheet | Price attached to the wrong product or variant | Rows wrap across lines; merged cells break row order |
| Financial summary | Forecast reported as actual | The qualifier sits in a footnote far from the number |
| Scanned contract or invoice | Digits misread (8 vs 3, 1 vs 7) | Text comes from OCR, not from a text layer |
| Long, image-heavy report | Sections ignored or summarized vaguely | Large files may not be fully analysed |
Units, periods and qualifiers are the real problem
Most “wrong numbers” are the right digits with the wrong meaning. “4.2” is useless without “$ millions”, “FY2025” and “unaudited”. In a PDF those pieces are often spread across a table header, a caption and a footnote, and extraction keeps the digits while losing the context.
- Scale: thousands vs millions set once in a header (“in $000s”).
- Period: column headers like “2025 / 2024” that disappear when a table is flattened.
- Currency: one symbol at the top of a column, none next to the values.
- Qualifiers: forecast, pro forma, unaudited, “up to”, excluding one-offs — usually in small print.
- Definitions: revenue vs GMV vs bookings, gross vs net.
What you can fix in the PDF yourself
- Export from the source application so text stays selectable; avoid scans and “print as image”.
- Put units and periods in each column header, not only in a page title.
- State key figures once in a sentence with their full context: “Revenue FY2025: $4.2M, unaudited.”
- Add data labels or a small value table under every chart.
- Avoid merged cells and tables that break across pages.
- If precision matters, share the spreadsheet or a CSV next to the PDF, as OpenAI recommends.
- When asking ChatGPT, point it to a specific page or section rather than the whole file; OpenAI suggests asking about sections or splitting large files.
Why fixing your own prompts is not enough
You can check ChatGPT’s answer for yourself. You cannot check the answers your readers get: an investor, a customer’s procurement team or a regulator each uploads the PDF to their own assistant, which extracts the numbers again, differently each time. An outdated copy keeps circulating after you publish a correction.
The durable fix is one canonical, text-based version of the document at a permanent address that every AI reads instead of guessing from the layout.
How The Papers solves it
The Papers turns a PDF (or PowerPoint, Excel, Word) into a fidelity-checked AI Version at a permanent URL: a web page plus Markdown and JSON. Before publishing, every number, currency, date and period is checked to match the source exactly, and no figure may appear that is not in the source. Qualifiers such as forecast, unaudited or pro forma stay attached, and each value keeps a reference to its source page.
Where the source is ambiguous — say, whether a figure is revenue or GMV — the question goes to you and your answer is recorded as owner-confirmed. The document gets an AI Readiness score; “Fidelity checked” requires zero unresolved critical issues. When you update the document, the link stays the same.
FAQ
- Is a scanned PDF worse than an exported one?
- Usually yes. A scan has no text layer, so values come from OCR, and OpenAI notes that scanned files and image-based tables may not be extracted reliably. Export from the original application when you can.
- Will a better model fix this?
- Partly. Better models read layouts better, but a PDF still does not state which period, currency or qualifier belongs to a number. Structured text removes the guesswork for every model.
- Should I just send the spreadsheet instead?
- For raw data, yes — OpenAI recommends spreadsheets or text files when exact values matter. But readers also need the narrative, and a spreadsheet alone loses it. A fidelity-checked AI Version keeps both.
- Does The Papers check that my numbers are correct?
- No. It checks that the AI Version is faithful to your source — the same figures, periods and qualifiers. It does not certify that the content itself is true.