Foreign archives still arrive as photographs of pages. A researcher who needs a Polish parish register, a stamped university transcript, or a signed exhibit cannot treat that file as a block of copy. The stamp, the column, and the signature are part of the claim. Generic translation starts by pulling strings out of the image. That is why desks that do scanned document translation have to keep the page, not just the words.
Citation work in our desk fails when a number leaves its cell. A grade on a transcript is only usable if it still sits beside the course title. A clause in a scanned contract is only usable if the stamp still sits on the same sheet. Extract the text, paste it into a chat box, and the spatial evidence is gone. The translation may be fluent. The source is no longer inspectable.
Extracted Text Drops The Evidence Researchers Need
Internet research specialists are trained to keep provenance. A screenshot without a URL is weak. A quote without a page is weaker. A scan is both image and document: columns, seals, handwritten notes in the margin, a notary stamp in the corner. OCR that only “gets the words out” turns that object into a paragraph. The paragraph cannot show where the stamp was.
A literature review that cites “page 4 of the scanned gazette” is making a location claim. If the translation lives in a text box with no page geometry, that citation becomes a guess. The reviewer who cannot match the sentence to the leaf will treat the note as unsourced. In a CIRS-style workflow, unsourced is the same as unused.
The same failure shows up in market and legal intake. A price in a scanned invoice is only a price while it sits in the amount column. Once it is a floating number in a paragraph, nobody can defend it against the original sheet. The reconstruction job exists to keep that defence.
That is not a preference about pretty layout. It is a verification problem. If a later reviewer asks “which box held the date,” a stripped text file cannot answer. The original scan could. Any process that throws the geometry away has already spent the evidence.
Convert Only Versus Convert And Translate
Scanned.to splits the job on the upload screen. Convert Only rebuilds an editable page from the image. Convert & Translate rebuilds the page and changes the language. Researchers should pick the mode that matches the question they are actually asking.
If the file is already in a language the team can read, conversion is enough. If the file is a German contract the UK team must understand this week, translation is the point. Mixing the two — converting, then pasting the new Word file into a second tool — starts a second geometry loss.
Researchers also mix up “searchable” with “translated.” A searchable PDF lets a query find a word on the page. It does not put that word into English. Convert Only is the right button when the language is already known and the problem is that the scan is an image. Convert & Translate is the right button when the language itself is the barrier. Pressing both paths on the same file, or exporting then re-uploading, is how stamps drift.
| Need | Mode | What to check after download |
| Read a scan in its original language | Convert Only | Tables and stamps still in place; text selectable |
| Read a foreign scan in English | Convert & Translate | Same geometry, language changed, dual view usable |
| Court-ready sworn wording | Neither mode | Send the scan to a certified translator |
The table is a routing rule. Most research intake belongs in the first two rows. The third row is a different profession. Do not ask an AI reconstruction to impersonate a certified translation.
How A Scan Becomes An Editable Page
The published path is three steps. Upload a clear scan or photo as PDF, JPG, or PNG. Choose the target language. Download a file that still carries images, signatures, and the original arrangement of blocks. The reconstruction keeps text regions separate from photos, then writes the new language back into the same hierarchy.
Output options include an editable Word file, a searchable PDF, or a shareable page. Dual view is the research-friendly one: original and translation side by side, so a date can be checked against the ink. Processing time scales with page count. That is a processing note, not a promise that a 40-page bundle will be ready before a meeting.
Upload The Scan Then Pick The Language
New accounts get 2 free trial pages. That is enough to test one certificate or the first two pages of a contract, not a whole archive. Pay-as-you-go pages do not expire. Monthly plan pages reset each billing cycle. If the project is irregular — one foreign PDF this month, nothing next month — the never-expire pile is the one that matches the calendar.
Keep the original image in the same folder as the output. Name both with the same source id. When a later editor asks where a number came from, open dual view and point at the box. That habit costs seconds and saves a rework loop in which someone re-types a date from memory.
Shareable pages are for a colleague who needs to look, not for the canonical file. The research drive should hold the downloaded Word or searchable PDF, plus the original scan. A link that expires with the upload window is not an archive.
Multi-page bundles hide the failure in the annex. Page one is often a clean cover. Page seven is the table with the stamp. Spend the trial pages on the annex. If those two sheets survive, the rest of the file is a volume problem. If they do not, more pages will only multiply a bad reconstruction.
Dirty Inputs That Break The Reconstruction
The page is only as good as the scan. A curved book page, a thumb over the stamp, or a dark shadow across a table will take labels with them. A flat, well-lit phone photo is easier to read than a shot taken at an angle. That still does not rescue a blurry image of a folded form.
Printed text is the input the engine is built for. Handwriting is a maybe. Margin notes in a 1920s register may come through; they may not. If the research claim depends on a handwritten addition, keep the original image next to any machine output and do not let the output stand alone.
Phone Photos Need Flat Light And Focus
Lay the page flat. Fill the frame. Avoid a glossy flash that whites out the stamp. If the date in the header cannot be read by eye on the photo, the reconstruction will not invent a trustworthy one. A warped photo of a folded letter merges columns. That output should be discarded. The source photo is the thing to fix, not the language setting.
A researcher who has to translate scanned documents for a literature review should run this check before paying for a long PDF. Two trial pages on the worst leaf in the bundle will tell you whether the rest is worth spending. If the worst leaf fails, split the file or rescan. Do not push a bad image through a higher page pack and hope volume helps.
Binders are worse than single sheets. A 20-page scan often includes a blank verso, a skewed photocopy, and one page shot at an angle. Process those as a stack and the good pages inherit the failures of the bad ones in review time, even if the engine handled them separately. Pull the worst leaf out. Test it. Then decide whether the binder is one job or five.
Color stamps and pale blue forms are another trap. Low-contrast ink that the eye can still parse may drop out of a reconstruction. If the dual view shows a blank where the original has a seal, that page would never clear legal as a stand-alone exhibit. Keep the photograph. Use the translation as a reading aid only.
Pages Credits And What The Trial Actually Gives
Billing is by page. A multi-page PDF spends one page per sheet. Lite lists a 100MB file cap; Pro lists 300MB. Those caps matter for fat image PDFs. They are not a reason to advertise a magic ceiling. If a scan is larger than the plan allows, split it. Do not compress until the stamp becomes mush.
The 2-page trial is the honest first hour. Use it on a page that contains a table and a stamp, not on a clean title sheet. Title sheets always look fine. The failure sits in the annex. A lost hour spent on a pretty cover page teaches nothing about the file you actually need.
Scanned.to also sells monthly volume for people who convert every week. Researchers with bursty archives are usually better on pages that never expire. The cost signal is calendar fit, not a cheaper headline rate.
Court Filings Still Need A Human Certificate
This is machine translation of a reconstructed page. It can support a business review, a personal reading, or an internal note. It is not a certified translation for court. Handwriting may fail. Stamps can sit in the right place and still be unread if the ink is faint. Keep a human in the loop whenever a filing, a visa, or a medical handover depends on the wording.
Keep The Page When The Source Is A Scan
Scanned.to fits research intake where the object is a photographed page and the team needs that page readable in another language without stripping the geometry. It is the wrong tool when the file is already born-digital Word, or when a regulator requires a sworn translator’s stamp on the wording itself.
Run the small test. Two trial pages. A table. A stamp. Dual view open. If the date still sits in the same box, the reconstruction did the research job. If it does not, rescan before buying more pages. The archive is the evidence. The translation is only useful while that evidence is still on the sheet. If the box moved, the citation moved with it. File the original scan beside the translation before anyone cites a number from a box.
