A table can contain every expected number and still say the wrong thing. Move a value one column to the right and it may describe another year, treatment, or unit. Reviewing scanned PDF table OCR means checking the relationships between labels and values, not just whether the digits look plausible.
Start with the original page image and one complete table. Inspect the header, body, notes, and any continuation before copying values into a spreadsheet or report.
Map merged headings before checking cells
A header may span several columns, with a second row supplying the individual labels. For example, “Annual production” might cover two years, each divided into separate product types. If the spanning header disappears, the exported rows can look tidy but lose their meaning.
Write the full label for a sampled value: table title, group heading, column heading, row heading, and unit. Then see whether a reader can reconstruct that label from the conversion alone.
Do the same for merged row labels. A category printed once beside several rows may govern all of them. A blank cell beneath it is not necessarily missing data, and filling every blank with zero would change the source.
For later Markdown representation choices, see PDF tables to Markdown. Here the question is whether the original table was recognized correctly.
Trace values in both directions
Pick a value in the scan and locate its row and column in the result. Then pick a value in the result and trace it back to the scan. The reverse check can reveal duplicated rows or values that have no visible source.
Sample different positions: first row, middle, last row, and a row containing a note marker or unusual punctuation. Check decimal separators, minus signs, thousands separators, percentages, ranges, and parentheses. These can affect meaning even when the surrounding digits are correct.
Preserve distinctions such as blank, dash, “not applicable,” and zero. The table's legend may assign different meanings to them. Do not normalize them for convenience until you understand that legend and record the change separately.
Read the units with the heading and notes
A value headed “Revenue” is incomplete if “in thousands” appears only above the table or in a note below it. Confirm that this qualifier survives and remains associated with the right values.
Inspect units that appear inside cells as well. Superscripts, prefixes, and degree symbols may be small enough to disappear in a low-quality scan. A change from milligrams to grams is not a minor typography issue.
If you later translate the document, separate unit recognition from language choice. A translator should not be relied on to discover that OCR lost an exponent. Review the recognized source first, then compare the translated headings and explanations.
Join multi-page tables cautiously
Look for “continued,” repeated column headings, split rows, and notes that appear only after the last page. A repeated header should not become an extra data row; a repeated row label may be a continuation rather than a second observation.
Check the last data row before the page break and the first after it. If a description wraps onto the next page, confirm that its values still belong to the same item. Footnotes at the end may apply to rows on earlier pages.
Keep page boundaries in your review record even if you remove them from the working table. They let you reconstruct where an omitted line or unexpected duplication came from.
Record discrepancies without guessing the numbers
Use a simple record: source page and table, row label, column label, disputed value, source reading, and correction status. Mark an unreadable value as unresolved. Do not choose a number simply because it makes the totals add up.
Totals and ranges can be useful cross-checks, but they are not proof. The original itself may contain rounding or editorial errors. Keep recognition corrections distinct from any later interpretation of the data.
PDF Craft recognizes tables during PDF conversion, including scanned material. Select Markdown if editing and reuse are the goal, then compare the downloaded result with the original. Review complex or faint tables carefully; no general promise covers every merged cell or cross-page structure.
Before using the table in research notes, verify its title, units, sampled rows, and notes together. The PDF-to-Markdown research workflow explains how to retain that source context when the table leaves the original document.