A PDF can look perfectly readable on screen and still give a translation tool nothing to translate. If you can’t select a sentence, search for a name, or copy a paragraph, you’re probably looking at page images rather than usable text.
The fix isn’t to keep trying different translation buttons. An image-only PDF needs an extra step: optical character recognition, or OCR. OCR reads the writing in an image and creates machine-readable text. You then check that text, translate it, and compare the translated file with the original pages.
That sounds straightforward until a seal is mistaken for a letter, a table turns into a jumble, or a translated paragraph no longer fits its box. Here’s how to choose a workflow that handles those risks instead of hiding them.
What you’re actually translating when a PDF has no selectable text¶
An image-only PDF is a file whose pages contain images of text rather than text characters a computer can select and search. A scanned paper document is a common example, but a PDF made from page photographs can have the same limitation.
Try selecting a few words and copying them into a plain-text editor. If nothing appears, or the copied text is missing or garbled, the file may have no usable text layer. A text layer is the machine-readable content attached to a page image. The page can look normal even when that layer is absent.
The distinction matters because translation software doesn’t translate what a person can visually read from a page. It needs access to the words as text, or a recognition system that can turn the visible words into text first. The Library of Congress notes that PDF/A files made from scanned page images at conformance level B don’t necessarily support text indexing or extraction for quotation. A readable-looking file isn’t proof that its words are machine-readable (Library of Congress format description).
OCR is a recognition step, not translation. The software examines the page image, identifies likely letters and words, and creates text that can be searched or copied. A translation tool can then process the recognized text. If OCR reads 0 as O, drops a minus sign, or joins two table columns, the translation may carry that error forward.
Google Translate Help says: “You can find text in images and scanned .pdf pages in the output document but they aren’t translated.” (Google Translate Help)
The practical point is that uploading a scan directly doesn’t guarantee that the text inside its images will be translated. Check whether the tool recognizes the page content, and don’t treat an output file as complete just because it opens successfully.
A mixed PDF can be especially confusing. Some pages may contain selectable text while others are scans. Google Cloud’s documentation says scanned content in a PDF that contains both scanned and native PDF content isn’t translated in that workflow (Google Cloud document translation documentation). A few searchable pages therefore don’t prove that the whole file is ready.
The workflow: recognize, check, translate, compare¶
A dependable image-PDF workflow separates four jobs: identify the page type, recognize the text, correct material errors, and review the translated output against the images. Each step catches a different kind of failure.
- Check the file before choosing a tool. Select and copy a passage from more than one page. Look for blank pages, skewed scans, photos, tables, stamps, and pages that behave differently from the rest. Mark which pages have selectable text and which don’t.
- Find an editable original if one exists. A DOCX or PPTX source usually gives a translation workflow more structure than a flattened PDF. Google Cloud recommends translating PDF content in DOCX or PPTX form before converting it back to PDF when that editable original is available, because layout and style are generally preserved better in those formats (Google Cloud documentation).
- Run OCR on scan pages. Set the recognition language where the tool asks for it. Recognition works by matching visual patterns to characters; it doesn’t know from appearance alone whether a blurred mark is a letter, a stamp, or a pen stroke.
- Review the recognized text against each page image. Correct names, document numbers, dates, negative signs, units, amounts, headings, and text inside tables. Fix errors that could change meaning before passing the content to a translation system.
- Translate the checked text. Use a workflow that keeps the text’s relationships to headings, labels, fields, and tables where possible. A plain stream of words may lose the structure that tells a reader which value belongs to which label.
- Compare the output with the original pages. Check for missing paragraphs, untranslated text, shifted labels, broken tables, text that overlaps or spills out of its area, and images or seals that are absent from the result.
- Save the source and working files separately. Keep the original scan unchanged, then retain the OCR text and translated output as separate working versions. That makes it easier to trace a disputed name or number back to the visible page.
The OCR correction step isn’t optional for every document, but skipping it is risky whenever a single character can change the meaning. A wrong date in a form, a misplaced decimal in an invoice, or a name read incorrectly can survive translation because the translator sees the OCR output, not the original ink.
NARA explains that OCR analyzes an image and creates a machine-readable text layer. Its guidance also says that OCR isn’t required for digitizing records (NARA transfer-guidance tables). That distinction is useful: OCR is a practical requirement for many translation workflows, but it isn’t a universal requirement for every archival scan.
For US federal records transferred to NARA, the original bitmap must remain intact: OCR must not replace it, alter visible content, or degrade the original image. NARA accepts uncorrected OCR when the image is preserved appropriately (NARA transfer-guidance tables). Those rules apply to that federal records-transfer context, not to every translation job. Still, preserving the source image is a sensible way to keep an independent reference while you correct recognition errors.
Your options: OCR software, cloud translation, or a document workflow¶
The right option depends on the file and the cost of a mistake. A simple printed page with clean contrast is a different job from a mixed scan with tables, handwriting, stamps, or legal wording.
| Option | What it does well | What you still need to check | Best fit |
|---|---|---|---|
| OCR tool, then translate | Separates text recognition from translation, so you can correct OCR before translating | Recognition accuracy, formatting, and whether the translation step keeps page structure | Scans where you want control over the recognized text |
| Consumer document translation | Convenient for supported document uploads | Whether text inside page images is actually translated; file limits and output layout | PDFs with machine-readable text, or quick drafts after OCR |
| Cloud document-translation API | Provides a document-oriented workflow and supports PDF translation | API setup, scanned-PDF formatting loss, page limits, mixed-content behavior | Teams building a document process into software |
| Document translation tool with OCR | Combines recognition and translation and may rebuild document structure | Handwriting, degraded scans, omissions, and whether the result suits formal use | Teams that want a formatted working file and a review step |
| Human review | Can resolve uncertain text and assess context | Time, cost, and whether the reviewer can inspect the original scan | High-stakes, unclear, or formal documents |
Google Translate’s browser document-upload feature accepts PDF, DOCX, PPTX, and XLSX files up to 10 MB, with PDFs limited to 300 pages, according to its Help page accessed April 11, 2026. The same page says document translation isn’t available on smaller screens or mobile, and it warns that text in images or scanned PDF pages may appear without being translated (Google Translate Help). Check the live limits before relying on them because service limits can change.
Direct upload is therefore a poor bet when the only words in the PDF are inside page images. If the document is already searchable, consumer document translation may be a convenient first draft. If it’s a scan, run OCR first or choose a workflow that explicitly recognizes scanned pages.
Google Cloud Translation is a separate API workflow, not the same thing as the consumer upload page. Its documentation supports PDF and DOCX document translation and describes retaining much of the source formatting and layout. For scanned PDFs, the same documentation cautions that formatting can be lost, especially in complex tables, multi-column pages, and graphs with labels or legends (Google Cloud documentation).
As documented on April 11, 2026, Google Cloud’s online document translation limits PDFs to 20 MB, and scanned PDFs to 20 pages. The 300-page online limit applies to native PDFs when the native-PDF setting is enabled (Google Cloud documentation). Those figures describe that service’s documented limits at that time; check the current documentation before building a workflow around them.
A document-translation tool can put OCR and translation in one workflow, but automation doesn’t remove the need to check a scan. For example, ChatsControl accepts scanned PDFs and photos, reads the page content, and returns a formatted Word file that follows the page layout. A typical use is turning a clean printed scan into an editable draft for review; handwriting and very degraded scans remain unreliable and need a human check. A Word result is also not the same as a finished, certified PDF.
For a team, the key comparison isn’t just “Which tool translates PDFs?” Ask what each option does with image-only text, whether you can correct OCR before translation, what file you receive, and how you can check layout and omissions. Those questions reveal whether a tool fits the actual job.
A practical comparison before you commit to a workflow¶
A useful comparison starts with the file’s structure, not a feature list. The table below summarizes what to verify in each route; service limits and behavior can change, so the cited documentation should be checked again before use.
| Question | Consumer document upload | Cloud document API | OCR plus separate translation |
|---|---|---|---|
| Does it accept a PDF? | Yes, within the documented upload limits | Yes, within the documented API limits | Depends on the chosen OCR and translation tools |
| Does upload alone translate image text? | Google says text in scanned pages may remain untranslated | Supports scanned PDFs, but mixed files have a specific limitation | OCR must create usable text before translation |
| What happens to layout? | Check the output page by page | Google says scanned PDFs can lose formatting | Depends on how you preserve and rebuild the structure |
| Can you correct OCR errors first? | Not necessarily in the upload workflow | Workflow-dependent | Yes, if you inspect and edit the recognized text |
| Is it suited to mobile browser use? | Google says document translation isn’t available on smaller screens or mobile | API integration rather than a consumer mobile upload page | Depends on the tools and devices you choose |
Google Cloud can detect a PDF’s source language when you don’t supply a source-language code, and its documentation says the detected code appears in the response (Google Cloud documentation). Automatic language detection can help when a file’s language is unclear, but it doesn’t validate OCR. A scan that reads the wrong characters can also lead a system toward the wrong language identification.
When should you stop automating and ask a person to review?¶
Ask for human review when the scan is hard to read, a handwritten note carries meaning, or an uncertain character affects a name, amount, date, legal condition, or medical detail. OCR can produce plausible-looking text from an unclear mark. Plausibility isn’t proof that the recognition is right.
A human reviewer needs the original image, not just the extracted text. The page image provides visual evidence for checking whether the OCR captured a stamp, a handwritten correction, or a character that the text layer doesn’t show. If an authority requires a certified translation, an automated draft alone may not meet that requirement.
For printed, legible documents, automated OCR can prepare a working copy efficiently, but keep the review proportional to the stakes. A brochure and a signed agreement don’t need the same level of scrutiny. In both cases, make sure someone checks what the system could not confidently read.
A page-by-page checklist for translation teams¶
Use this checklist before you send a translated scan to a colleague, client, or reviewer. It covers the steps where image-PDF workflows most often lose information: recognition, structure, and comparison with the source.
Before OCR¶
- Keep the original file untouched. Make a working copy so image processing or text correction can’t replace the source you need for comparison.
- Test more than one page. A document can mix born-digital text with scanned pages; check representative pages rather than assuming the entire PDF behaves the same way.
- Locate an editable source. If the original DOCX or PPTX exists, consider using it instead of reconstructing the document from a PDF.
- Note page features that matter. Mark tables, columns, stamps, seals, signatures, form fields, and small print so the reviewer knows what to inspect.
During OCR¶
- Set the recognition language where available. A mismatch between the page language and the recognition setting can make similar-looking characters harder to distinguish.
- Watch the page image while correcting text. Don’t approve a name, amount, date, or identifier just because the OCR output looks fluent.
- Check reading order. Two columns read as one continuous line can scramble the relationship between headings and text.
- Treat handwriting as uncertain. Recognition of handwriting and very poor scans is unreliable in the ChatsControl workflow; the product’s quality grade can flag scan quality, but it doesn’t replace checking against the original.
After translation¶
- Compare every page with its translated counterpart. Look for omitted lines, untranslated labels, lost footnotes, missing stamps, and text that no longer fits.
- Recheck values independently. Compare numbers, dates, names, units, and references against the image rather than relying on the translated text alone.
- Inspect tables and graphs closely. Google Cloud warns that scanned-PDF translation can lose formatting in complex tables, multi-column layouts, and graphs with labels or legends (Google Cloud documentation).
- Confirm the intended output format. A Word file can make editing easier, but a recipient may require a PDF or a certified translation. Verify the recipient’s actual requirements rather than assuming the format is interchangeable.
- Record unresolved doubts. Flag unclear words and keep the source page reference beside each question. A reviewer can then return to the exact image instead of guessing from context.
A built-in review view can help when the main risk is missing or altered text. In ChatsControl, the bilingual side-by-side view displays the source and translated text together, with translated text highlighted and the original available on hover. That can make post-editing more direct, but the source image still matters when OCR itself is uncertain.
Pitfalls that create a polished but unreliable translation¶
The most dangerous output isn’t always an obviously broken file. A translation can look clean and still omit text, attach a value to the wrong field, or carry an OCR error into a sentence that reads naturally.
Mistaking a readable page for searchable text. A page that looks sharp to a person may have no usable text layer. The Library of Congress’s description of scanned PDF/A files explains why visual readability doesn’t guarantee text extraction (Library of Congress format description). Test selection and copying before you choose a translation method.
Assuming a translation upload includes OCR. Google Translate Help explicitly warns that text inside scanned PDF pages can appear in the output without being translated (Google Translate Help). Search the output for the source-language text or compare each page image with the translation.
Treating OCR as a fact-checker. OCR converts visual patterns into likely text; it doesn’t know which reading makes sense in a contract, invoice, or certificate. The correction workflow described in OCR software documentation highlights uncertain words so a user can compare them with the page image before accepting them. Correct material recognition errors before translation, not after the error has been rewritten in another language.
Flattening the page and losing relationships. A form isn’t just a sequence of words. A label and value belong together because of their positions. Two columns, table cells, or graph labels can become ambiguous when a workflow extracts text without preserving those relationships. Review layout as information, not decoration.
Overwriting the scan with corrected text. Some records workflows require the original image to remain unchanged. NARA’s instructions for permanent records transferred to the US National Archives say OCR must not alter or replace the original bitmap (NARA transfer-guidance tables). That rule is specific to federal records transfer, but keeping the image intact gives any translation workflow a reliable reference.
Using scan settings without considering the job. NARA’s archival guidance specifies 300-600 ppi for clean, high-contrast bitonal text; 300-400 ppi grayscale for poor-legibility text; and 300-400 ppi color when color matters. Those are specifications for scanned textual records transferred to NARA, not universal minimums for every translation job (NARA scanned textual records guidance). Don’t treat an archival transfer specification as a guarantee of OCR quality.
Sending a scan without checking the expected file. A team may need editable Word text for review, while a recipient may need a final PDF or a certified translation. Confirm the deliverable before the work starts. When the document is official, a machine-translated draft and a certified translation are different things.
Google Cloud’s documentation states: “Translating scanned PDF files results in some formatting loss.” (Google Cloud document translation documentation)
That warning is a reason to inspect the output, not a reason to assume every page will fail. Complex tables, columns, and graphs deserve particular attention because their visual structure carries meaning that plain text can lose.
FAQ¶
How do you translate a scanned PDF that has no selectable text?¶
Run OCR to convert the page images into machine-readable text, correct important recognition errors, and then translate that text. Compare the translated output with the original page images to catch omissions and layout problems.
Can Google Translate translate text inside scanned PDF pages?¶
Google Translate Help says that text in images and scanned PDF pages may appear in the output without being translated (Google Translate Help). Run OCR first or use a workflow that explicitly processes image text.
Should I run OCR before translating an image-only PDF?¶
Yes. OCR recognizes the words in the page image and creates text for a translation tool to process. Check the recognition against the scan before translating if names, dates, amounts, or other details matter.
How can I check whether OCR has recognized a scanned PDF correctly?¶
Compare the OCR text with the corresponding page image. Focus on names, numbers, dates, stamps, tables, small print, and low-contrast text; correct meaningful errors before translation.
How do I preserve the layout of a translated scanned PDF?¶
Use an editable DOCX or PPTX source if one exists. Otherwise, compare each translated page with the scan and inspect tables, columns, graphs, labels, and text placement; Google Cloud warns that scanned-PDF translation can lose formatting (Google Cloud documentation).
Can I translate a scanned PDF on a phone?¶
Google Translate’s browser document-translation feature isn’t available on smaller screens or mobile, according to its Help page accessed April 11, 2026 (Google Translate Help). Other tools have their own device support, so check the workflow you plan to use.
Can I use OCR without changing the original scan?¶
Yes. Keep the original page image as the reference and use recognized text as a separate text layer or working copy. NARA’s guidance for federal records transferred to the US National Archives requires preserving the original bitmap, although that rule isn’t a universal requirement for every translation job (NARA transfer-guidance tables).