PDF to Word Guide: Review Text, OCR and Document Structure
Admin ·
PDF to Word reads selected PDF pages in your browser, lets you correct the extracted content, and creates an editable DOCX. Start with a short representative page range. Choose existing text, OCR for printed scans, or optional structure reconstruction, then compare the review with the original before downloading.
A PDF describes fixed pages; an editable Word document uses paragraphs, tables and other structures that can reflow. Conversion is therefore a reconstruction task, not a guarantee that every line, font or page break will match. Keep the source PDF and review the resulting Word file in the application where you will use it.
Begin with a page that reveals likely problems
Choose one PDF or drop it into the workspace. The default is Automatic reading, English OCR and plain editable paragraphs. Enter a short range in Pages, provide the known password if the PDF requires it, and choose Read PDF text. Selecting a file alone does not create a Word download.
Use a sample with the features you need: ordinary text, a scan, a table, or two columns. If tables, headings, lists or figures matter, enable Reconstruct document structure before reading. That changes both the review and the default output choice. Inspect the first result before attempting a larger selection.
- Keep the original PDF as the visual reference.
- Choose the printed OCR language deliberately; changing the interface language does not select its OCR model.
- Check names, numbers, missing lines and reading order before exporting.
Microsoft Support: opening PDFs in Word and conversion limitations
Choose how each selected page should be read
Automatic uses an existing text layer whenever its extracted text is nonblank after trimming. It uses OCR only for pages without that text. It does not measure whether the existing layer is complete or correct. A heading above a scanned invoice can therefore cause Automatic to return the heading while missing the invoice text inside the image.
Choose OCR every selected page when the useful words are in pictures or the text layer is unusable. This rasterizes the selected pages and recognizes their printed text, including existing visible text. Recognition can introduce errors, so compare the result again rather than assuming that OCR is an improvement for every document.
| Reading mode | Useful starting point | Main limitation |
|---|---|---|
| Automatic | A document containing ordinary text pages and separate scanned pages | A partially useful text layer can prevent OCR on a mixed page |
| Existing PDF text only | Selectable text that you want to inspect without OCR | Text in pictures and form field values can be absent |
| OCR every selected page | Printed scans, mixed pages or an unusable text layer | At most 10 OCR pages; recognition must be checked |
Adobe Acrobat learning: scanned paper and OCRTesseract documentation: improving recognition qualityTesseract.js: image recognition and PDF support boundary
Use page ranges and check their order
An empty Pages field selects the entire source. Enter 1-3, 5 to read pages one through three followed by page five. Individual ranges must ascend, but separate segments keep the order you entered. For example, 3,1,3 produces three review entries: page three, page one and page three again. Repeated pages count toward selection and OCR limits.
The source can contain at most 300 pages, with no more than 50 selected pages in a job. Selecting one page from a 301-page source does not bypass the source limit. Split that source into suitable PDFs first. For mixed-language scans, read separate ranges with the appropriate OCR language rather than expecting one selection to detect every language.
Match the printed language and prepare readable scans
The OCR choices are English, French, German, Italian, Portuguese, Spanish, Simplified Chinese, Hindi, Arabic and Bengali. One chosen model applies to all OCR pages in that reading job. Indonesian and Urdu interfaces do not add Indonesian or Urdu recognition models. Handwriting and languages outside this list are unsupported here, and OCR does not translate text.
Clear, upright pages with readable printing give a better starting point than blurred, skewed or low-resolution scans. The tool renders OCR pages at 200 pixels per PDF inch; it has no user resolution or scan-enhancement control. Structure OCR uses unrotated word boxes with automatic deskew disabled. Correct the source orientation before importing when necessary. A high confidence estimate can still accompany omitted lines or wrong identifiers.
Tesseract documentation: improving recognition qualityTesseract.js: image recognition and PDF support boundary
Correct the content against every original page
The review shows the original page beside editable text and provides a page selector. Check proper names, dates, decimal separators, account numbers, punctuation and complete lines. Keep identifiers such as 9007199254740993 as text; the converter is not asking you to round them or treat them as spreadsheet numbers.
In plain mode, edit the text directly. Line breaks become Word paragraphs. You can also type into an empty review when text-only reading found no existing text, but a blank plain-text review cannot produce a Word document. If you need missing printed content from a scan, read it with OCR and review the new extraction. Reading again clears the previous review, so download important edits first.
Microsoft Support: opening PDFs in Word and conversion limitationsTesseract documentation: improving recognition quality
Reconstruct blocks when editable layout is useful
Structure mode attempts to recover paragraphs, headings, supported lists, tables, one- or two-column regions and figures. Usable PDF tags can guide that process; other content is grouped from positions and appearance. Neither tags nor geometric guesses establish the correct reading order automatically. Compare blocks with the source, and move, edit or remove them as needed.
Table cells remain editable, including manually added tables. Check column alignment, merged headers and the order in which cells are read. Lists on different source pages receive separate Word numbering definitions, so review continuation and restarts. A document can look plausible while putting a right-hand column before its left-hand text.
| Review target | Useful correction | Check after export |
|---|---|---|
| Headings and lists | Choose heading levels, list style and numbering; reorder blocks | Heading hierarchy and list continuation |
| Tables | Correct cell text, rows, columns and merged-cell interpretation | Each value belongs under the intended header |
| Columns and figures | Review region order and add or remove crops | Reading order, missing artwork and duplicated text |
Microsoft Support: opening PDFs in Word and conversion limitationsPDF.js API: document loading, text and page rendering
Treat fonts, colors and figure crops as approximations
Structure mode can retain matched font family names, sizes and supported solid text colors, with controls for review. It does not embed original fonts. Your browser or word processor may substitute an unavailable family. Transparency, patterns, strokes and clipping are not reproduced as an exact PDF appearance; heed the displayed notes and compare the preview.
Figures are rendered crops, not the original embedded image files. Native structure uses a 150-PPI page render and can omit full-page images or images overlapping text to avoid duplication. Vector artwork, masks and repeated or grouped images may need a manual crop. A crop can contain text pixels already present in editable blocks, so inspect both for duplication. OCR structure does not automatically turn every picture into a separate figure.
Microsoft Support: opening PDFs in Word and conversion limitationsPDF.js API: document loading, text and page rendering
Keep plain text and reconstructed edits distinct
When a structure review exists, Word content lets you choose Reviewed structure or Plain editable paragraphs. These versions retain separate edits. Changing a table in structure does not also rewrite the plain-text version; changing plain text does not update the structured blocks. Choose the intended version before downloading DOCX or reviewed text.
Changing reading settings, choosing another file or reading again clears the review. Changing only Word output settings keeps the reviewed content and removes the old result so that it can be generated again. If an export error was displayed, changing those output settings also clears that old error; the next export still has to complete successfully.
Choose page size, create DOCX and inspect it in Word
Plain output offers US Letter or A4 and creates editable paragraphs without recreating the PDF’s original tables, images, links, forms, fonts or layout. Structured output uses the reviewed blocks and offers source page dimensions as well. Enabling reconstruction initially selects structured output and source dimensions; switching to plain text changes a source-size choice to Letter.
The page-break option starts each source page on a new Word page, but cannot guarantee identical pagination within it. Text, margins, substituted fonts and your word processor can all change reflow. Choose Create Word document, wait for the result, and download the DOCX. Structure reconstruction itself requires source pages from 0.5 to 22 inches per side. For pages outside that range, turn reconstruction off and read again for plain Letter or A4 output, or prepare a standard-sized PDF first; the other input and rendering limits still apply.
Microsoft Support: opening PDFs in Word and conversion limitations
Check file, selection and review limits separately
The 50 MB input allowance means 50,000,000 bytes. A PDF must also meet the page and rendering limits; a small compressed file can still contain a very large page. Selecting fewer pages can reduce cumulative work but does not bypass the total source-page or file-byte limit. Structure and OCR use larger renders than a plain-text preview.
Text limits count JavaScript UTF-16 units rather than visible characters: some symbols use two units, and combining sequences can use several. The displayed review count excludes separators inserted between pages for TXT and Word Counter. Structured export also has bounds on blocks, styled runs, figures and table size. When a limit is reached, reduce the selection or simplify the reviewed document.
| Scope | Implemented allowance | What to do when it is exceeded |
|---|---|---|
| Source PDF | 50,000,000 bytes and 300 pages | Split or prepare a smaller source first |
| Selected pages | 50 entries; at most 10 needing OCR | Choose a smaller range and avoid unintended duplicates |
| Reviewed/exported text | 1,000,000 UTF-16 units; plain DOCX at most 20,000 paragraphs | Export smaller portions |
| Rendered pages | 8192 pixels per side; 8,000,000 pixels per page; 80,000,000 total | Use a smaller range or standard-sized source pages |
| Initial review assets | 50,000,000 bytes across previews, source renders and figures | Reduce page selection or use plain text |
| Structured DOCX | 2000 blocks, 50,000 styled runs, 100 figures and 50,000,000 figure bytes; output at most 100,000,000 bytes | Simplify blocks and figures or split the job |
| Each structured table | 200 rows, 12 columns and 2000 cells | Divide a large table into suitable parts |
| Job deadlines | 180 seconds for reading; 60 seconds for Word creation | Retry a smaller job on an available device |
PDF.js API: document loading, text and page renderingTesseract.js: image recognition and PDF support boundary
Use the failure message to choose the next step
A .pdf filename does not guarantee readable PDF contents. This workflow checks the file and uses strict PDF diagnostics, with a limited allowance for repairable font hints. It can reject a damaged file that another viewer displays with repairs. A successful preview elsewhere does not establish that its text or structure is safe to reconstruct here.
If reading fails or is cancelled, the interface does not offer a partial set of reviewed pages. Fix the input or options and read again. If Word creation fails, your review remains available for correction or a TXT download. An unavailable local engine or model may require the relevant site assets to load successfully before retrying.
| Symptom | Likely check | Next step |
|---|---|---|
| No text from a scan | Text-only mode cannot read picture pixels | Choose Automatic or OCR every selected page |
| Only the heading was extracted | A mixed page has some selectable text | Try forced OCR and compare the whole page |
| Wrong or missing words | Language, scan quality or recognition limitations | Select the right model and correct against the source |
| Page or rendering limit | Source size, selected count or raster dimensions | Prepare a smaller source or range |
| Layout differs in DOCX | Reconstruction, font substitution or Word reflow | Review blocks and inspect the exported document |
PDF.js API: document loading, text and page renderingTesseract documentation: improving recognition quality
Use the known password and keep edits before rereading
For a protected PDF, enter its correct password in the local password field and try reading. This is not password recovery. The field clears after every read attempt, including failure and cancellation, so another attempt may require entering it again. The chosen PDF remains selected after cancellation.
Cancel stops the current read or export. Clear file removes the selected source, password, range and review, while other reading choices and Word settings can remain. Check those settings before a different document. Download edits you need before clearing, rereading or leaving the page; an on-screen review is not a saved document.
Download TXT or send the reviewed text to Word Counter
Download reviewed text uses the output representation currently selected under Word content. For structure, that means text from the reviewed blocks, including table text and applicable list markers; figure pixels contribute no text. The TXT contains all review pages in selection order, separated by two line breaks.
Count reviewed text sends that same text to Word Counter in the current tab. Its 1,000,000-unit allowance includes the inserted page separators, so a review at its own text limit can be too large to transfer. Download smaller sections if needed. The transfer is temporary and is not stored in a URL or browser storage; Clear in Word Counter removes the transferred text and undo history.
Distinguish local document work from site downloads
The PDF, password, text review and Word creation are processed on your device. The workflow does not upload those document bytes for conversion. Loading the website still requires page and script requests, and PDF support files or the selected OCR model can be fetched when needed. Site analytics may also make requests. Local processing is not a guarantee of fully offline first use or zero network activity.
Existing-text reading does not need an OCR model. When OCR is needed, the chosen language model loads with the OCR engine; it is not an automatic download of every model. A failed asset download can prevent recognition even though document processing is local. Download the finished files to retain them independently of the tab.
Tesseract.js: image recognition and PDF support boundaryPDF.js API: document loading, text and page rendering
Use a short acceptance check before sharing the result
For an invoice, compare the customer name, invoice identifier, line items, quantities, decimal separators and total with the PDF. For a report, also check heading levels, list continuation, column order, table headers and figure captions. Recognition confidence is an engine estimate, not a measured accuracy score for your document.
Open the downloaded DOCX in your intended editor and check the pages at a readable zoom. Correct any reflow there or adjust the source review and export again. Keep the PDF alongside the editable version when its original visual arrangement matters. For a fixed visual copy, retaining the original PDF may be more useful than trying to reproduce every detail in Word.
Microsoft Support: opening PDFs in Word and conversion limitationsTesseract documentation: improving recognition quality
Common questions
How do I convert a PDF to editable Word?
Choose one PDF, select a reading mode and page range, then choose Read PDF text. Compare and edit the review, choose the Word output settings, create the document and download its DOCX result.
Can this read scanned PDF pages?
Yes, for supported printed languages within the OCR limits. Automatic uses OCR only where no nonblank text layer is extracted; choose OCR every selected page when a mixed page or unusable text layer needs recognition.
Why did Automatic miss text inside a picture?
Automatic checks whether existing extracted text is nonblank, not whether it covers the page. A selectable heading can prevent OCR of the rest. Forced OCR may recover the pictured words, but its result needs checking.
Does changing the interface language choose the OCR language?
No. The initial OCR model is English. Choose the printed language explicitly. Ten models are available; Indonesian and Urdu are interface languages without corresponding OCR models here.
Does OCR translate text or support handwriting?
It recognizes supported printed text and does not translate it. Handwriting and languages outside the offered model list are unsupported.
Will my Word file look exactly like the PDF?
No. Plain output contains editable paragraphs. Structured output reconstructs supported blocks and figures, but font substitution, uncertain grouping and Word reflow can change the result. Check the downloaded file.
Can I edit tables, headings and columns?
Enable reconstruction before reading, then inspect the detected blocks. Review heading levels, list numbering, table cells, merged headers, column order and figures. Detection is approximate even when PDF tags are available.
Do edits in plain text also change the structured version?
No. The two representations keep separate edits. Word content selects which version supplies DOCX, TXT and the text sent to Word Counter.
How many pages can I convert at once?
The source may contain at most 300 pages. A job may select at most 50 entries, with no more than 10 pages requiring OCR. Repeated selections count again, and file, text, rendering and time limits also apply.
Does choosing a short range bypass a large source limit?
It reduces selected work but does not bypass the 50,000,000-byte or 300-page source limit. Split an oversized source into suitable PDFs before selecting it.
Can I convert a PDF that needs a password?
Enter the correct known password locally before reading. The field is cleared after every attempt. This tool does not recover or bypass an unknown password.
Why is Word Counter disabled for a full-length review?
Its limit includes the two line breaks inserted between review pages. Those separators can push an otherwise allowed review past 1,000,000 UTF-16 units. Download and count smaller portions.
Are the PDF or password uploaded for conversion?
This conversion workflow processes them locally on your device. The site can still request scripts, PDF support assets, the chosen OCR model and analytics. Local conversion does not imply zero network requests.
What should I keep before cancelling or clearing?
Download reviewed text or the completed DOCX if you need to retain edits. Cancellation keeps the selected PDF, while Clear file removes the source and review. Reading again resets the previous review.