ChatGPT got numbers wrong in my PDF

If ChatGPT got PDF numbers wrong, put a value under the wrong table column, dropped a negative sign, changed units, or mixed up rows and columns, check the extraction before prompting again.

Practical guideReviewed Aug 2026
Short answer

Separate extraction from analysis. Reproduce the target table with its original headers, units, signs, and page reference; check the extraction; only then ask AI to analyze it. If the PDF is scanned, run OCR first and inspect the recognized values.

Why readable PDFs still produce wrong values

A PDF stores instructions for drawing a page, not necessarily a clean grid of rows and columns. Text can be extracted in an order that differs from what your eyes see. A value may remain visible while its relationship to a year, header, subtotal, or unit is lost.

Common silent errors
  • A value shifts into the wrong table column after an empty cell collapses.
  • Rows or columns are mixed up.
  • A negative sign or accounting parentheses disappear.
  • Units such as thousands, millions, %, or currency detach from the value.
  • OCR substitutes 0/O, 1/l/I, 5/S, or 8/B.

Why correct-looking PDF numbers land in the wrong table column

PDF extraction can preserve the digits while losing the layout that tells you which row, column, year, or unit they belong to. Empty cells, wrapped labels, repeated headers, and page breaks can shift relationships without making the output look obviously broken.

Reproduce the table before analyzing it. Preserve blank cells, original headers, units, decimal points, negative signs, and page references. Mark any uncertain cell as unclear instead of inferring a value.

OCR failure and table failure are different

OCR converts page images into text. It may be necessary for a scan, but accurate characters do not guarantee accurate table structure. A digitally generated PDF can have excellent selectable text and still lose row-to-column relationships.

Use the selectable-text test as a first clue. If words cannot usually be selected, prioritize OCR. If text is selectable but values are wrong, prioritize structured table extraction.

If a table looks correct but copied numbers do not, compare a PDF page with its extracted text before analysis.

A safer workflow for PDF tables

  1. 1
    Isolate the table.

    Work with the smallest relevant page range.

  2. 2
    Reproduce it before interpreting it.

    Preserve row labels, headers, units, decimals, and signs.

  3. 3
    Check the extraction.

    Spot-check dense rows and every consequential value.

  4. 4
    Analyze only after you’ve checked the extracted table.

    Keep facts and interpretation separate.

A safer starting instruction

“Before analyzing, reproduce the relevant table as structured rows with the original labels, units, signs, and page number. Mark any ambiguous cell as unclear. Do not infer missing values. Wait for me to confirm the extraction before interpreting it.”

Before you trust the next result

  • Confirm the tool can identify actual content from the PDF.
  • Request page or section support for important claims.
  • Compare exact values, dates, clauses, and chart details with the original.
  • Treat unclear or missing source evidence as unresolved — not as permission to guess.

Your failure may involve more than one cause.

Use the six-question diagnostic to decide whether OCR, table structure, or visual layout should be repaired before analysis.

Diagnose why your PDF numbers went wrong