A counterparty sends the engagement letter as a PDF and asks for two changes. A portal rejects PDFs and demands JPGs. You need three paragraphs of a filing quoted in a letter, and copy-pasting out of the viewer mangles every line break. Same kind of file, three different jobs — and the route that solves one makes the other two harder.
Underneath all three sits the same trap: every route behaves differently depending on whether the file was authored on a computer or scanned from paper, and only one of them tells you which kind you are holding. One of the others charges you, then hands back a file that opens in Word, is named .docx, and holds a picture of the page with not one editable word in it. It looks exactly like a success.
If the PDF was authored digitally — you can drag-select the words — and you need to edit it, convert it to Word, then check the layout before you rely on it. If you need the words and not the layout, extract the text instead: cheaper, exact rather than reconstructed, and it tells you immediately if the file is a scan. If a portal only accepts images, render the pages. If the PDF is a scan, none of these gives you editable text — converting it produces a Word file with a picture inside. Upload it to Filly as a form, where OCR runs, or use a dedicated OCR product. And if the real job is "convert every PDF that lands in this folder", Filly has no public API; that is a script's work, not this toolbox's.
Four routes, two kinds of PDF
| Route | From a digitally authored PDF | From a scan | Does it warn you? | Cost |
|---|---|---|---|---|
| PDF → Word | An editable .docx, layout rebuilt — close on plain text, drifting on tables | A .docx containing a picture of the page. Nothing in it is editable | No. It downloads normally and looks like it worked | 2 credits |
| PDF → plain text | The exact text, line breaks kept, pages labelled | Nothing — so it refuses, and says the file is most likely a scan | Yes, by name, and the credit is refunded | 1 credit |
| PDF → JPG/PNG | A picture of each page. The text layer is gone | A re-render of the same image, no more readable than the scan | No. A page renders whether or not there is text under it | 1 credit |
| Retype it by hand | Faster than any of the above for a paragraph or two | Exact, and often the only honest answer for one page | — | Free |
First, work out which PDF you have
Ten seconds, before you spend anything: open the file and try to drag-select a word. If the words highlight, there is a text layer and every route below is available. If dragging draws a rectangle instead, it is a scan — an image of a document — and only OCR will get the words back. That one test decides the rest; for the longer diagnosis, why a PDF won't let you type covers the flat-versus-scanned distinction in full.
When you need to edit it: PDF to Word
This is the route people reach for first, and it is right when the job really is editing. What matters is that conversion is a reconstruction, not a translation. Nothing in the file says "two-column layout with a nested table"; the converter infers structure from where each character sits and rebuilds something similar. Plain text comes across cleanly; multi-column pages, intricate tables and unusual fonts are where it drifts. Microsoft says the same of Word's own built-in PDF conversion — free if you already have Word — noting it works best with PDFs that are mostly text, and that the converted document may not correspond page-for-page with the original because lines and pages break in different places.
So open the result before you send it anywhere. In Filly this costs two credits rather than one, because converting is one of the operations that spends money outside our own servers: it runs through an external conversion service billed per file. Extracting text and rendering images run locally, which is why those cost one.
When you need the words, not the layout: PDF to text
Most of the time nobody wants the layout. They want three paragraphs pasted into a letter, or a document they can search. Extracting the text is a different operation from converting, and a better-behaved one: it reads the text layer already in the file, so the output is exact rather than inferred — line breaks preserved, each page labelled in a multi-page file, and a plain .txt that find-in-file can search. Structure is what you give up: tables and columns come out as sequential lines, and extraction is capped at 30 pages.
Its most useful property is one nothing else here has: it refuses to fail silently. When a PDF has no text layer the tool stops, says the file is most likely a scan, and refunds the credit rather than handing back an empty .txt that reads like a bug. Because the refund lands exactly when it refuses, learning you are holding a scan costs nothing — which makes this the diagnostic to run first, before two credits tell you the expensive way.
When something else needs an image: PDF to images
Plenty of agency and employer portals accept JPG and PNG and reject PDFs outright, and being right about that changes nothing. Rendering the pages is the answer there, and only there. Pick PNG for text and forms and JPEG for mostly photographic pages. Rendering is fixed at roughly 144 DPI rather than configurable and capped at 30 pages; a one-page PDF comes back as a single image, anything longer as a .zip with the pages named in order. It is a one-way door: an image is a picture of the page, so whatever text layer the PDF had is absent from the output — the entire point when a portal wants an image, and a loss you notice far too late if it wasn't.
The scan, and why it defeats three of the four routes
A scanned document has no text underneath: nothing to extract, nothing for a converter to rebuild a layout from. Text extraction says so. Image rendering hands back a re-render of the picture it started with, at a fixed resolution that may be lower than the original scan. Conversion is the one that hurts — two credits for a Word file containing an image, which is why our PDF-to-Word page says so in its limits.
OCR is a genuinely different operation — reading letters out of pixels — and in Filly it does not live in the conversion tools at all. It lives on the form-upload path: upload the document as a form, and OCR runs in your browser, in English, on the pages that look scanned, feeding the recognised text into placement and filling. Filling a paper form from a photo walks that route. If you want a searchable document rather than a filled form, that is someone else's job: Google Drive will OCR a PDF into a Doc free, within its documented constraints — the file 2 MB or smaller, the page right side up, and lists, tables, columns, footnotes and endnotes described by Google as not likely to be detected.
The round trip, and when not to take it
The return leg is Word to PDF, and it exists because a .docx re-flows on the recipient's machine while a PDF does not. It is also a reconstruction: the service renders the document with its own layout engine and substitutes any font it does not have. Microsoft's own guidance on font embedding describes the mechanism from the other side — embed the fonts and the layout and styling do not change on a machine that lacks them. A substituted font has different character widths, so lines wrap in new places and page breaks move.
Hence one hard rule, worth more than the rest of this section: never round-trip a document that has already been signed, or one whose pagination is cited somewhere. A filing that refers to "page 12" is describing pagination the round trip is free to change. And a round trip does not carry a signature across intact: a drawn one comes back as an image rebuilt twice, and a cryptographic one does not survive the conversion at all. Check the page breaks before filing.
Frequently asked questions
How do I convert a scanned PDF to editable text?
Not with a conversion tool — a scan has no text to convert, so what comes back is your image inside a Word file. You need OCR: in Filly, upload the document as a form rather than running a conversion.
Why did the text extraction say my PDF has no text?
Because it is a scan — an image of a page with nothing underneath. The tool stops deliberately and refunds the credit rather than returning an empty .txt that looks like a bug.
Should I pick PNG or JPEG?
PNG for anything with text — forms, letters, filings — because it is lossless, so small type stays legible with no artefacts around thin strokes. JPEG for mostly photographic pages, where the size saving is large and artefacts do not matter.
Will my formatting survive a PDF-to-Word conversion?
Usually for straightforward documents, often not for complicated ones. Plain text and simple tables come across; multi-column layouts, intricate tables and unusual fonts are where the reconstruction shows. Open the file and check before you send it.
Why do the conversions cost two credits and the extractions one?
Because they are paid for differently. Extracting text and rendering images run on our own servers, so they cost one credit. Both conversion directions run through an external service billed per file, so they cost two — the same reason Filly's two AI-backed tools, redact personal details and AI page organise, also cost two. Subscribers have all of them included; the free plan starts with ten credits on signup and ten fills a month, and a failed run is refunded.
Can I do this automatically for every file in a folder?
No. Filly has no public API: every toolbox operation is a signed-in, manual upload, capped at 25 MB per request, with a 30-page ceiling on the text and image tools. Watching a folder is a script's job. Filly's case is the other one — a document on your desk, and a decision about which route to take before you spend a credit on the wrong one.