Convert a PDF table to Excel

A PDF does not contain a table. It contains glyphs at coordinates, and what you see as a grid is an arrangement that happens to line up. Every tool that gets a spreadsheet out of a PDF is inferring that arrangement back, and the honest thing to do is say so rather than imply the structure was in the file all along.

This works it out from geometry. Text sharing a baseline is a row. Left edges that recur down the page are a column. Both tolerances come from the size of the text itself, so a table set in seven-point type and one set in fourteen are read the same way — a fixed threshold in points would find one and miss the other.

What it leaves out, and why that is the point

A page whose rows are mostly incomplete is not a table. It is prose, a cover sheet, a signature block, an address panel. Feeding it through the same clustering produces columns of sentence fragments, and one such page in a twelve-page statement makes the whole workbook untrustworthy — because you cannot tell by looking which rows came from the table and which came from the paragraph above it.

So those pages are left out, and the result tells you how many pages had a table in them. Ten of twelve is a useful number. A workbook that quietly included the other two would not be.

Where it will be wrong

Three things need a model of the page rather than a clustering of it, and this does not attempt them.

A merged cell spanning two columns lands in one of them. A cell whose contents wrap onto a second line becomes two rows. Rotated text is read on the baseline the rotation gives it, which is rarely the one you meant.

Each of those would be guessable, and guessing produces a spreadsheet that looks right and is quietly wrong somewhere in the middle. A sheet that is obviously missing something is a better outcome than one that is subtly incorrect, because you can see the first one.

What it will not do is silently drop a value. Two cells landing in the same column are joined rather than one being discarded, so a row that came out wrong is visibly wrong rather than short.

Continuation pages

A table running across pages is one table. Every page contributes its rows, and a page that repeats the column headings — which most multi-page tables do — contributes everything below them rather than adding the headings again as data.

You get one worksheet, not one per page. Twelve pages of a bank statement come out as twelve pages of rows in a single sheet, which is what anybody exporting a statement actually wants.

It needs real text

A scan is a picture of a table. There are no glyphs at coordinates, only pixels, so there is nothing to cluster and nothing will be found. If selecting the text in your own PDF reader does not work, this will not work either — run the file through OCR first to give it a text layer, then convert.

Nothing leaves the device

The document is read and the workbook is written by a worker inside your browser. There is no upload, no account and no server copy, which matters when the table you are extracting is a payroll, a client list or a set of financials.