Drop in a text-based PDF and get its tables as a spreadsheet, one sheet per page. Worth knowing what is actually happening: a PDF contains no tables at all, only text fragments with coordinates — the table is something your eye assembles from the alignment. So this reconstructs it, grouping fragments that share a baseline into rows and clustering their left edges into columns. That works well on a clean report and less well on a page of prose, so each page is scored on how table-like it looks and flagged when it is not. If your PDF is a scan, there is no text to take positions from at all, and you will be told to run it through OCR first rather than handed an empty sheet. Two sliders let you tune the detection when a row or column is being split or merged.
| Layout | Result |
|---|---|
| Clean aligned columns | Accurate |
| Merged header cells | Header may land in one column |
| Multi-line cells | Each line becomes its own row |
| Prose paragraphs | Flagged as low confidence |
| Scanned image pages | Detected — OCR needed first |
No. Extraction happens in your browser. That matters here, since the PDFs people convert to spreadsheets are usually statements and financial reports.
Because a scan is a picture of a page, with no text in it. There are no coordinates to work from — there is nothing but an image. Run it through the PDF OCR tool first to produce real text, then bring the result back here.
Detection works from where text starts, so an unusual layout can confuse it. If one column is being split in two, raise the column gap. If two are being merged, lower it. If a single row is coming out as two, raise the row tolerance. Drop the file in again after changing either.
That the page does not look much like a table — usually because there is only one column, or the column count changes from row to row, which is what prose produces. The data is still exported, but expect text split at arbitrary points rather than real columns.
Because a multi-page report usually has a different table on each page, often with different columns. Merging them into one sheet would misalign everything. Sheets keep them separate, and combining them afterwards is easy if they do match.
They are written as the PDF stores them, which is text. Excel will usually recognise plain numbers on import; currency symbols and thousands separators often need a column format applied afterwards.
PDFs do not contain tables. They contain text fragments with coordinates, and the table you see is something your eye assembles from the alignment — so extraction means reconstructing that structure from positions. Rows come from grouping fragments that share a baseline, with a tolerance rather than exact equality, because a bold header sitting beside plain text is rarely on precisely the same y and exact matching splits every such row in two. Each fragment is compared against its row's first item rather than its last, which sounds like a detail and is not: comparing to the last lets a row drift downward a tolerance at a time and swallow the row beneath it. Columns come from clustering the left edges of every fragment across all rows — in a real table each column's cells start at the same x, so the starts pile into as many clusters as there are columns. Splitting each row on its own whitespace instead, which is the obvious approach, produces a different column count per row and aligns nothing. Each page is checked for how table-like the result is, from the column count and how consistent it is across rows, and a page that looks like prose is flagged rather than silently exported as nonsense. A PDF with no text at all is a scan — images of pages with nothing to take positions from — and is called out as needing OCR first rather than producing an empty spreadsheet. Everything runs in your browser.