What this tool does
Upload a picture of a page and the text appears below it, in the order a person would read it. That covers a photo of a book, a screenshot of an app that will not let you select text, a scanned contract, a whiteboard after a meeting, and a page of your own handwriting.
Each page comes back separately, with a confidence mark, so on a batch you can see which page needs a second look.
Nothing you upload becomes a public page. The file is deleted as soon as the job finishes, and the text stays attached to your job for you to copy or download.
What it reads well, and what it does not
Reliable: printed pages, receipts, forms, screenshots, slides, book pages photographed flat, and neat handwriting in the writer's usual hand.
Usually fine: cursive that a stranger could read, a whiteboard photographed square-on, a scan with a coffee ring on it, and mixed pages where a table sits in the middle of prose.
Genuinely hard, and it will say so rather than invent an answer: personal shorthand, faded thermal receipts, text at a steep angle, dense small-type columns, and anything photographed while moving.
Unreadable words come back as [illegible]. That marker is the point: a reader who cannot tell which words were guessed has to check every one.
Photographing a page so it works
Accuracy is decided before you upload anything, by how the picture was taken. Four things matter, in this order.
- Fill the frame with the page. Text that is 40 pixels tall reads perfectly; text that is 8 pixels tall does not, whatever the megapixel count of the camera.
- Shoot square-on. A page photographed from an angle has converging lines and uneven focus, and the far end of every line is the part that fails.
- Find even light. One window beats one lamp; a shadow across the middle of a page costs you that band of text. Turn the flash off.
- Hold still, or rest the phone on something. Motion blur cannot be recovered by any reader.
There is no step five. Resolution beyond about 2000 pixels on the long edge adds upload time and nothing else, which is why oversized images are scaled down here before they are read.
Tables, screenshots and PDFs
Switch "Read as" to tables and every grid on the page comes back as rows and columns instead of a run of words. That is the difference between retyping a supplier list and downloading it, and it is why the spreadsheet export exists: one sheet per table, with numbers stored as numbers so a column of prices adds up without being cleaned first. When the table is the whole point, Image to Excel is built for exactly that and adds a CSV download.
Screenshots are the easiest input there is, because the text was rendered rather than photographed. Error messages, chat threads, dashboards and anything inside an app that blocks copying all read cleanly.
PDFs are handled two ways. A PDF exported from Word, Pages or a website already contains its text, so it is pulled straight out of the file — no model involved, the result is exact rather than approximate, and those pages are free because nothing had to be recognised. A PDF that is really a scan is read page by page like any other image, and charged like one. You do not have to know which kind you have; the file is checked and the cheaper path is taken automatically.
How the free tier works
Three pages a day, across up to three uploads, cost nothing. You see the text, the tables and the confidence marks, and you can copy all of it. There is no account and no email.
Upload more pages than your allowance covers and the pages inside it are read in full, the rest are listed as unread, and the result says exactly where it stopped. Nothing is charged. On a scanned PDF, only the pages you are entitled to are ever sent to be read, so a preview costs the same whether the file has four pages or four hundred.
Beyond that it is 1 credit for each page that has to be recognised. Credits are bought once, never expire, and work across every tool here: a 30-page scanned report costs 30 credits, roughly $1.30 from the middle pack. The Word and spreadsheet downloads cost 3 credits each, once per job — re-downloading the same file later is free.
A PDF that already carries its own text is the exception. Extracting it costs nothing on our side, so it costs nothing on yours: up to 40 pages a job, free, without touching your daily allowance.
Checking the output before you use it
Read the numbers first. Text errors are obvious on sight; a transposed digit in a total or a date is not, and those are exactly the values people copy without looking. Then check proper nouns, which have no context to be corrected from, and anything you intend to quote.
That is a minute of work per page — the same minute a data-entry service builds into its price. What you are buying here is the retyping, not the checking.
Example result
Two photos of a handwritten stock list, plus a scanned 4-page supplier PDF
6 pages read, in order, each with a confidence mark. Tables: 3 grids lifted out, 41 rows total. Downloads: XLSX for the stock list, DOCX for the supplier terms.
Cost: 3 free pages, then 3 credits for the rest, plus 3 for the spreadsheet — about $0.30.
How this compares
Where another tool is the better choice, it says so.
Google Lens / Google Drive OCR
BETTER THEREFree at any volume, built into a phone camera, and excellent on printed text in good light.
BETTER HERELens gives you a block of text with no page structure and no table extraction, and Drive's OCR ignores handwriting. Here a table stays a table, multi-page batches stay in order, and handwriting is read rather than skipped.
Adobe Acrobat OCR
BETTER THEREWrites the text back into the PDF as a searchable layer while keeping the original page exactly as it looks, which is the right tool for archiving.
BETTER HEREIt is a subscription, it is a desktop app, and it is built for scans of printed documents. This reads a photo of handwriting on a phone in ten seconds for a few cents, with no install.
Free OCR websites (Tesseract-based)
BETTER THERENo credits at all, and on a clean 300 dpi scan of printed text the accuracy is competitive.
BETTER HERETesseract has no idea what a word means, so it fails on handwriting, curved pages, screenshots with UI chrome and anything low-contrast. It also has no notion of a table. This is a different class of model, and it tells you when it is unsure.
Your questions, answered.
- Is it really free?
- Three pages a day, including tables and the confidence marks, with no account and no email. Credits cover anything beyond that plus the Word and spreadsheet downloads.
- Can it read handwriting?
- Yes, and it is the main reason to use a model-based reader rather than a classical OCR engine. Neat handwriting in a familiar hand comes out close to perfect; personal shorthand and doctors' notes come back with [illegible] markers where the words genuinely cannot be read.
- Can I upload several images at once?
- Up to ten files in one job. Each image and each PDF page counts as one page against your allowance, and the output keeps them separate and in order.
- Does it keep the layout?
- It keeps reading order, line breaks, lists and tables. It does not reproduce fonts, columns or positions — if you need the layout preserved exactly, what you want is document translation or a PDF editor, not OCR.
- What about a scanned PDF with 200 pages?
- Only the pages your allowance or balance covers are read, and only those pages are ever sent to be read. A job is capped at 40 pages so one upload cannot run away with your balance; run the rest as a second job.
- Why was my PDF free?
- Because it already contained its text, so nothing had to be recognised — those pages are extracted locally, charged nothing and not counted against your daily allowance. Only pages that exist as pixels cost a credit.
- Which languages work?
- Most languages written in Latin, Cyrillic, Greek, Arabic, Hebrew, Devanagari, Chinese, Japanese and Korean scripts. Set the language when a page mixes two and you only want one of them; otherwise leave it on auto.
- Do you keep my images?
- The upload is deleted as soon as the job finishes. The extracted text stays attached to your job so you can come back to it, and is never rendered on a public page or used to train anything.
- Is it accurate enough for accounts or medical records?
- Not without checking it. Treat the output as a first pass that removes the retyping: verify every figure you will act on. For anything legally binding you need a human transcription with a certificate of accuracy.
- What does [illegible] mean in my text?
- That a word was unreadable and no guess was made. It is deliberate: a marker you can search for is worth more than a plausible invention you would never notice.