PDF is a container rather than a format, and two invoices that look identical on screen can behave completely differently when something tries to read them. One carries a text layer produced by the supplier's system. The other is a photograph or scan wrapped in a PDF. This single distinction predicts most of your extraction quality.
How to tell them apart
Try to select the text on screen. If you can, the document carries a text layer and extraction reads exact characters, including amounts and references. If you cannot, it is an image and needs optical recognition, with accuracy depending on resolution, contrast and how straight the page was when it was captured.
Why the difference is large
A text-layer PDF gives exact characters, so figures and references are read rather than inferred. An image introduces a recognition step whose errors concentrate precisely where you least want them: digits in amounts, and references mixing letters and numbers. Everything downstream inherits whichever of the two you fed in.
What to do about it
Ask suppliers to email system-generated PDFs rather than scans, which most larger ones can do because their systems produce them anyway. It is a cheap request with a real effect on accuracy, and it usually outperforms changing extraction products, which is the more expensive way to address the same symptom.
And handle the image ones deliberately
Consistent scanning at a sensible resolution, capture at arrival, and confidence thresholds that route doubtful fields to a person. Where documents arrive as photographs from sites, an app that rejects a poor image at the point of capture is worth more than any downstream improvement.
Questions people ask about pdf invoice data extraction
Does a text layer guarantee correct extraction?
It removes the recognition error. Deciding which text is the total and which line belongs where is still interpretation, and line items remain the hard part regardless.
What about password-protected PDFs?
They block automated capture. Ask the supplier to stop sending them, or handle them as a named manual exception rather than letting them fail silently in a queue.
Can we convert images to text-layer PDFs?
Recognition can add a text layer, which helps searching. It does not recover information the image never carried clearly, so input quality still governs.