Extract a table from a PDF with ChatGPT and keep each value in its column

Give ChatGPT the PDF and the exact table you need, then ask for a literal table extraction before requesting analysis. Check the row labels, column headers, units, and footnotes against the PDF. A result can contain all the right numbers and still attach them to the wrong month.

The first goal is a checked copy of the table. Calculations and charts come after its structure and values are correct.

Identify the table and preserve its context

Attach a permitted PDF where file uploads are available, and identify the page and table title. OpenAI's file guidance, checked October 24, 2026, describes extracting and reorganizing information from attached documents. File access and features depend on your account and workspace.

Include the table's caption, headers, units, and footnotes in the requested extraction. A heading such as “amounts in thousands” changes every value beneath it. A footnote may explain why a total does not equal the visible rows.

If ChatGPT cannot read the table from the attachment, provide a clear image of that page through a supported image input. In ChatGPT web, OpenAI documents attaching, pasting, or dragging an image into the composer. Image-input instructions

Keep the edges of the table and its explanatory notes visible. Enlarging an unreadable image does not recover detail that was never captured. Obtain a better source when important characters remain unclear.

Request transcription before cleanup

Extract [TABLE TITLE] on [PAGE] from the attached PDF.
Return a Markdown table first. Preserve the row order, column order,
headers, units, signs, decimal places, blank cells, and footnote marks.
Repeat the relevant caption, unit note, and footnotes below the table.

Do not calculate, summarize, fill blanks, combine duplicate-looking
rows, or turn dashes into zero. Use [unreadable] for an unclear value
and list its row and column so I can check the original.
Identify continuation pages or repeated page headers separately.
After I verify the extraction, I may ask for a CSV version.

For a table continuing over several pages, give the page range and ask for the original row locations. A repeated column heading belongs in the header, not as a data row. A subtotal may need to remain a labeled row rather than being added to the records it summarizes.

Check a missing value before exporting

Here is an original fictional table for practicing the review. It is not a real PDF extraction or observed ChatGPT response.

Caption: Equipment loans. Unit: kits. A dash means not reported.

Location April May Two-month reported total
East 12 8 20
West — 5 5

East has a reported total of 20: 12 plus 8. West's reported total is 5, but that does not establish that April was zero. The source defines the dash as not reported.

Check that 8 remains under May for East and that the West dash survives extraction. If the output converts it to 0, correct the transcription before asking for a chart. A later calculation cannot recover the distinction after it has been erased.

Also compare the first and last data rows, any negative values, and each subtotal. If the table has merged headers, ensure every resulting column has an unambiguous name. “2026 / May / Loans” is different from a second unlabeled “May” column for another year.

Convert the checked table into a spreadsheet format

After reviewing the table, ask for CSV with a stated rule for missing values:

Convert the verified table below to CSV. Preserve all values as text
where needed. Use an empty field for not-reported values; do not use
zero. Keep the missing-value definition in a separate note.
Use a header row. Quote fields containing commas or quotation marks.
Do not include the caption or footnotes as data rows.
Verified table: [PASTE THE CHECKED TABLE]

For identifiers with leading zeros, import the relevant spreadsheet column as text. Open the exported result and compare its row count, headers, and a few distinctive values with the checked table. Your spreadsheet's automatic date or number conversion can change a value even after a correct extraction.

OpenAI's file review documentation recommends opening generated files in the appropriate viewer and giving focused revision instructions. If the task returns a downloadable spreadsheet, inspect that actual file rather than relying on the table shown in the chat.

Keep the PDF page reference and missing-value note with the extracted data. When you are ready to calculate, the data-analysis guide covers checking row definitions and denominators. For this table, preserve West's unreported April figure as missing until the source owner supplies it.

Keep reading