Scanned vs native PDFs: how to tell

A native PDF — one exported directly from Word, LaTeX, a browser or an accounting system — stores text as text: character codes, an embedded font, and a position for each run. Zoom to 800 % and the letters stay crisp, because they are being drawn from outline data at whatever size you ask for.

A scanned PDF is a photograph in a wrapper. Each page is a single raster image — often a JPEG or a CCITT Group 4 bitmap — and the PDF exists only to hold it. No character codes, no fonts, no coordinates. Everything a person reads on that page, the file itself knows nothing about.

Try to select a line of text with the mouse. If a tidy selection follows the words, there is a text layer. If the whole page highlights as one block or nothing highlights at all, it is an image. Then search for a word you can see with Ctrl+F: no match on a word that is plainly visible is conclusive.

Two more signals. Zoom deep: a scan turns to mush, a native page stays sharp. And look at the size — a text-only page is usually tens of kilobytes, while a colour scan at 300 dpi is commonly one to three megabytes per page, so a 20-page file of 40 MB is a scan regardless of what it looks like.

In TableExtractor the check is done for you: pages with no text runs are flagged in the page rail, and asking for a table on such a page returns a plain message instead of an empty grid.

Many scans have been through optical character recognition, which adds an invisible text layer positioned over the picture — rendering mode 3, text that is drawn but not shown. The page still looks like a scan and now behaves like a document: you can select, search and extract. These are called searchable image PDFs, and they are what a modern office scanner produces by default.

Extraction works on them, with a caveat: every recognition error is now baked into the data. In numeric tables the usual suspects are 0 read as O, 1 as l or I, 5 as S, and — most costly of all — a comma read as a full stop inside an amount. Geometry can be perfect and the value still wrong, which is why a total check matters more on OCR documents than anywhere else.

Offline options are the safe ones. OCRmyPDF is free, open source and adds a text layer to an existing PDF without touching the image; it uses the Tesseract engine and runs from the command line on Windows, macOS and Linux. Adobe Acrobat Pro and ABBYY FineReader do the same with a graphical interface, and FineReader is noticeably better on dense tables and on Cyrillic.

Free online OCR sites are convenient and they work by uploading your document to someone else's computer. For a printed article that is a non-issue. For a payslip, a medical result, a contract, a bank statement or anything with a name and an account number on it, read the retention policy first — and if the document would embarrass you in a leak, do it offline instead.

Scan quality decides the outcome more than the engine does. Three hundred dpi in greyscale, straightened, with the page flat and evenly lit, produces dramatically better numbers than a phone photo at an angle. If you are the one scanning, that choice is free.

A PDF is almost always the last step of a pipeline, not the first. The statement was generated from a database, the report from a spreadsheet, the price list from an inventory system. Asking the sender for the CSV, the XLSX or the export they already have takes one message and removes every problem in this article at once.

When that is not possible — a public register, an old archive, a regulator's publication — check whether the same body publishes machine-readable versions elsewhere. Statistical offices, central banks and open-data portals very often publish both, and only the PDF is linked from the press page.