The .pdf file.
- What it stores
- Glyphs at coordinates, not cells
- Two kinds
- Native, with a text layer, and scanned, which is an image
- Table structure
- Not stored. Inferred from geometry.
- Read here as
- Extracted for native PDFs, interpreted for scans
- Checked against
- Subtotals printed in the document, on scanned pages only
- Lost on the way in
- Payment references and long identifiers truncated by the layout
- Status
- Ready
- Fidelity
- Data extracted
- Largest file
- 25 MB free, 250 MB once a project is bought
What it actually is
Checked for a text layer first. A PDF written by software keeps its table structure; a scan is read as an image and told apart from the first kind.
PDF is a page description language. It records that a glyph appears at a position in a font at a size. Where you see a column, the file holds a set of numbers that happen to share an x coordinate, where you see a row, a set that happen to share a y coordinate. The grid exists in your head.
This is why extraction quality varies so much between documents that look equally tidy. A generator that emits cells left to right in reading order produces something almost trivially reconstructible. One that emits a whole column at a time, or that draws the rules as separate vector paths, does not.
Then there is the second kind of PDF entirely: a scan. That file has no text at all, only an image of one, and every figure has to be read visually. Selling both as "supported" without distinguishing them is how a product ends up quietly wrong about a number.
Worth knowing: A scanned PDF is read by eye, so it lands in the interpreted tier and is checked against its own totals.
Upload one
The analysis runs before there is anything to pay for.
You see what it found, and the evidence behind it, first.
No account needed to start. You only pay when you like what you see.
What goes wrong, and what is done about it
3 named failures, each with the repair.
A table split across a page break
The header repeats at the top of each page, the last row of one page and the first of the next are often halves of the same record, and a naive extraction produces one table per page with a header row buried in the middle of the data.
What is done: Nothing is stitched across the break. Each page becomes its own table, keyed by its page number, so a repeated header stays at the top of every page's table and a record split across the boundary stays in two halves.
Numbers that reconcile to nothing
A misread minus sign, a thousands separator taken as a decimal point, or a column that lost its rightmost digit to a clipped bounding box all produce a spreadsheet that looks perfectly clean and does not add up.
What is done: That check runs on scans, where every figure was read by eye: the rows are summed and compared against the subtotal printed in the same image. A native PDF's text layer is taken as written and is not reconciled against the document's own totals.
A scan that looks native
Many scanners run OCR and embed the result as an invisible text layer. The file then reports that it has text, and the text is whatever the OCR thought it saw, with none of the confidence information that would let you distrust it.
What is done: Nothing here tells the two apart. The only test is whether any text is present at all, so a page carrying an OCR layer is treated as native and that layer is read as written. Where the system that printed the document can export CSV or XLSX, that is the way past it.
If you have the choice, send something else
If the system that produced the PDF can also export CSV, XLSX or OFX, use that. A PDF has already thrown away the fields that were not printed: a bank statement PDF, for example, no longer has the payment reference that invoice matching runs on.
Questions people ask
About this format, not about the product.
Will it work on a scanned document?
Yes, and less well, which it tells you. A native PDF extracts close to exactly. A scan is read visually, lands in the interpreted tier, and every figure is re-derived against the totals printed in the same image before it is shown.
What happens to figures it cannot read?
They are listed as unreadable with their page number. Nothing is guessed to fill a gap, because a plausible invented figure is worse than an obvious hole.
Can it do a hundred monthly reports at once?
Zip them and upload the archive. Each one is read independently. Nothing worked out on the first is carried to the second, and a scanned page costs a model call of its own, so a hundred reports cost a hundred times what one does.
Keep reading
The jobs people do with this file, and the formats beside it.