# Correct OCR Text in a PDF: Fixing a Wrong Hidden Word

- Author: Schema Law (https://www.schemalaw.com/)
- Published: 2026-06-18T01:20:12Z
- Updated: 2026-10-11T19:22:20Z
- Category: Searchable PDFs
- Canonical: https://www.schemalaw.com/blog/correct-ocr-text-in-a-pdf

> You can correct OCR text in a PDF when your editor lets you change recognized words. Which errors to fix, the procedure, the log, and a fallback list.

You can correct OCR text in a PDF if your PDF editor has a tool for changing recognized words while leaving the scanned picture in place. Some editors have one and limit which words it reaches. Where yours does not, the choices are to run OCR again or to keep a corrections list beside the file.

The error you are fixing is invisible. The page shows "Okonkwo" because the page is a picture. The hidden text says "Okonkwa", so a search for the right name skips the page. Correcting it changes what search and copy return, and should change nothing a reader sees.

Put the corrections on lines you will search again. Fix the plaintiff and the defendant in the caption, the docket number, and a Bates number. If a cause of action line is misread, fix that too. After each change, search the word and compare the hit with the page image so the picture and the text name the same line.

## Can you correct the OCR text inside a searchable PDF?

You can correct the OCR text inside a searchable PDF in some editors and not in others. The hidden text is ordinary PDF text drawn invisibly, so nothing in the format forbids changing it. You need an editor that exposes recognized words for correction, and editors differ in whether they do and in which words they let you reach.

Before you spend time on it, settle three things about your own editor from its documentation:

- **Does it have a correction tool for recognized text at all?** This is separate from its general "edit PDF" tool.
- **Which words does the tool show you?** Some show only the words the program itself doubted. A confident misreading, which is the dangerous kind, may not be offered.
- **Does the general edit tool convert the page?** An editor that turns a scan into editable type has replaced the picture of the page with something new. That is a different document.

### One documented example

Adobe's help page on [fixing text recognition errors in scanned PDFs](https://helpx.adobe.com/acrobat/desktop/create-documents/scan-documents-to-pdfs/fix-scanned-text.html) describes Acrobat's version. It says: "If a match is uncertain, it marks the word as a suspect, visually showing the original image but placing searchable text behind it." The correction tool on that page walks you through those suspects. Adobe's page does not say whether a word Acrobat was confident about, and got wrong, can be corrected through the same tool, so test that on a copy of your own file.

Adobe's separate page on [editing scanned PDFs](https://helpx.adobe.com/acrobat/desktop/create-documents/scan-documents-to-pdfs/edit-scans.html) describes the general Edit tool differently: "Acrobat automatically applies OCR to your document and converts it to an editable PDF copy", and typed changes match "the original font in your scanned PDF". That is editing the visible document. It suits a form you are allowed to revise. It is the wrong tool for a scanned exhibit whose appearance has to stay as received.

### What correcting does not do

Correcting a hidden word does not change the page picture, does not make the scan more legible, and does not fix the same misreading on other pages. Each occurrence is its own word in its own place. And a corrected layer is still a working aid: the picture is the record of what the document says.

## Which wrong words in a text layer are worth correcting?

Correct the words someone will search for or copy: names of people and companies, amounts, dates, account and policy numbers, defined terms, and section references. Leave ordinary words alone. A layer that reads "tbe" for "the" loses you nothing, because nobody searches a contract for "the".

Use this to decide where the time goes:

| Kind of word | Why it matters | Correct it? |
| --- | --- | --- |
| Party, witness and company names | The first thing anyone searches; one wrong letter hides the page | Yes, every occurrence you can find |
| Dollar amounts and quantities | Copied into briefs, schedules and demand letters | Yes, and check the digits against the picture twice |
| Dates | Used to build chronologies | Yes |
| Account, claim, policy and case numbers | Searched exactly as printed; no context helps the software | Yes |
| Defined terms ("Premises", "Effective Date") | Searched to trace how a term is used through a document | Yes where the term is defined and where it does real work |
| Section and exhibit references | Used to jump around a long file | If you rely on them for navigation |
| Bates or page numbers | Used to cite | Yes if you cite from the layer; otherwise read them from the picture |
| Common words, boilerplate, letterhead | Not searched | No |

### Finding the errors in the first place

You cannot fix what you have not seen, and a hidden layer hides its mistakes. Copy a page's text into a plain editor and read it against the picture. The post on [how to see the text layer in a PDF](https://www.schemalaw.com/blog/how-to-see-the-text-layer-in-a-pdf) covers the methods for one page and for a whole file.

For names, work from a list. Write down every person and company in the matter, search the file for each, and compare the hit counts with what you expect from reading. A name that should appear on thirty pages and is found on eleven has been misread somewhere. The post on [searching a scanned PDF for a name](https://www.schemalaw.com/blog/search-a-scanned-pdf-for-a-name) covers the variant spellings to try.

### When the answer is "none of them"

If the file is one you will read once and never search, do not correct anything. If a page has errors in every line, word-by-word repair is the wrong job, and the question becomes whether to run recognition again.

## Correcting recognized text in a PDF editor without touching the page picture

Correcting recognized text follows the same procedure in any PDF editor that supports it: work on a copy, open the tool that reviews recognized text, change the word in the tool's own field while the scan stays on screen, save, and then prove by searching that the fix took and the page looks the same.

### The steps as Adobe documents them for Acrobat

Adobe's page on [fixing text recognition errors](https://helpx.adobe.com/acrobat/desktop/create-documents/scan-documents-to-pdfs/fix-scanned-text.html) gives this sequence:

1. Open the scanned file and select All tools from the global bar.
2. Select Scan & OCR from the All tools menu.
3. Select Correct recognized text from the left pane.
4. In the pop-up box, select the Review recognized text checkbox. Adobe says Acrobat then "displays the scanned image and the recognized text" and that "Unclear words are highlighted with boxes in the document."
5. Select a highlighted box and make the correction in the Recognized as field.
6. Compare the image with the text and select Accept. Acrobat moves to the next word.

Menu names move between versions, so follow the page for the version you have. If you use another editor, look in its documentation for a feature described as correcting, verifying or reviewing OCR results, and confirm that it changes the recognized text and not the visible page.

### What to do around those steps

The editor's steps are the middle of the job. The parts that protect you come before and after.

**Before:**

- Copy the file. Correct the copy. The file as received stays where it is.
- Write down the errors you intend to fix: page, the wrong text, the right text. You will need the list afterwards.
- Search for the wrong text and note the hit count.

**After:**

- Search for the corrected word. The count should have gone up by the number of fixes.
- Search for the wrong text. The count should have gone down by the same number.
- Select the corrected word on the page, copy it, and paste it into a blank document. It should paste as the right word.
- Put the corrected copy and the original side by side at 200% on each page you touched. Nothing visible should differ. If letters have changed shape, sharpened, or shifted, you edited the page, not the layer. Discard that copy and start again with the right tool.

### A worked example

A scanned guaranty names the guarantor as "Marisol Okonkwo" on nine pages. Search finds "Okonkwo" on six. Copying the text from the three missing pages shows "Okonkwa" twice and "0konkwo", with a zero, once. The paralegal, Teodor, records the three errors, opens the review tool on a copy, and finds two of the three offered as doubtful words. He corrects both. The third is not offered, because the software was sure of it.

He does not force the third with the general edit tool, since the guaranty is an exhibit. He adds it to a corrections list for the file, searches again, gets eight of nine, and notes the ninth page number so the next person knows where the gap is.

## Running OCR again versus correcting the text layer by hand

Correct by hand when a handful of important words are wrong on otherwise good pages. Run OCR again when errors are spread through a page or a file. Hand correction costs time per word. A fresh run costs one pass and a new round of checking, and replaces every word, including ones that were right.

| Situation | Cheaper repair |
| --- | --- |
| Three wrong names in a clean 40-page agreement | Correct by hand |
| One page reads as nonsense, the rest are fine | Re-run that page, or the file |
| Errors in every paragraph | Re-run with a different tool or a better scan |
| Text copies in the wrong order across columns | Re-run; this is a layout problem, not a spelling one |
| The layer came from a copier at scan time and nobody has checked it | Re-run and test before correcting anything |
| The paper is available and the scan is poor | Rescan first, then run OCR |
| A word sits under a stamp or signature | Neither is likely to help; record it on a corrections list |

Two cautions apply to a fresh run. First, a new run can get right what the old one got wrong and get wrong what the old one got right, so the tests have to be repeated from the start. Second, an existing layer has to be replaced, not added to, or the file ends up carrying two sets of hidden text. The post on [replacing a bad text layer by running OCR again](https://www.schemalaw.com/blog/convert-pdf-to-readable-text#replacing-a-bad-text-layer-by-running-ocr-again) covers how that is done.

If you run the file through Schema Law for a second reading, check the result the same way. The homepage says recognition is strongest on clean print and weaker on faint faxes, stamps over text and handwriting, and that nothing marks a passage as uncertain, so names, dates and amounts still need comparing with the original. The [common questions on the homepage](https://www.schemalaw.com/#questions) set that out.

A middle route exists for a long file: re-run the worst pages, keep the rest, and hand-correct the names that still fail. Record which pages got which treatment.

## Should a corrected text layer be logged?

Yes. A corrected layer is text that a person typed into a copy of a document, and anyone who later searches or quotes from that copy should be able to find out which words those were. A short log, kept with the file, answers the question before it is asked.

The log needs one line per change:

| Page | Text as recognized | Corrected to | Checked against page image by | Date |
| --- | --- | --- | --- | --- |
| 4 | Okonkwa | Okonkwo | T. Varga | 9 Oct 2026 |
| 7 | Okonkwa | Okonkwo | T. Varga | 9 Oct 2026 |
| 12 | $48,l50.00 | $48,150.00 | T. Varga | 9 Oct 2026 |

Add a header naming the file the corrections were made to, the file it was copied from, and the tool used.

### Why bother for a hidden layer

- **Quotes come from the layer.** When someone [copies a passage into a brief](https://www.schemalaw.com/blog/how-to-quote-text-from-a-scanned-pdf), the characters they paste are the layer's. If a figure in that passage was typed by a colleague, the person citing it should know to look at the picture.
- **Corrections can be wrong too.** A person fixing "48,l50" can type "48,510". The log lets a second person re-check exactly the words that were touched.
- **Copies travel.** If the corrected copy is sent to a client, an expert or another party, you may be asked how it differs from the original. Whether and how to disclose that is a judgment for the lawyer on the matter under the rules that apply to it. The log makes an accurate answer possible.
- **Files get regenerated.** If someone runs OCR on the original again next year, the hand corrections are gone. The log lets them be reapplied from a list instead of found again.

Name the corrected copy so it cannot be mistaken for the original, for example `Guaranty_2019_ocr_corrected.pdf`, and store the log beside it under a matching name.

## A corrections list for files whose text layer you cannot edit

When the layer cannot be edited, leave it alone and keep a corrections list: a small table that maps what the layer says to what the page says. Anyone searching the file reads the list first and searches for the wrong spellings as well as the right ones.

You will need this fallback when your editor has no correction tool, when the tool does not reach the word, when the file carries a certificate signature or a restriction that should not be disturbed, or when the file must stay exactly as produced.

### What goes in the list

```
File: Brandt_Production_Vol3.pdf (as produced, not modified)
Prepared by: T. Varga, 9 Oct 2026

Page   Layer reads        Page image reads     Note
0114   Okonkwa            Okonkwo              guarantor name
0114   0konkwo            Okonkwo              leading zero for O
0231   $48,l50.00         $48,150.00           lower-case L for 1
0388   (no text)          handwritten note     not recognized
```

Three habits make the list useful:

- **Search terms first.** Put the wrong strings where a searcher will see them, so "Okonkwa" and "0konkwo" get searched along with the real name.
- **Record gaps as well as errors.** A page with no text at all, or a handwritten note the software skipped, belongs on the list. It tells the reader that an empty search result does not cover that page.
- **Keep it findable.** Same folder, same base name, with `_corrections` on the end. A list nobody finds might as well not exist.

### Other ways to carry the correction

A comment or sticky note placed on the page can hold the right spelling, and a PDF program may be able to include comments in a search. Adobe's page on [searching PDFs in Acrobat](https://helpx.adobe.com/acrobat/using/searching-pdfs.html) lists "Include comments" as a Find option, for example. Check your own program's search options before relying on that, and remember that a comment is an addition to the file, which matters for anything that must stay unaltered. A corrected Word or plain text export, kept as a separate working transcript, is another option when people need to copy long passages. Schema Law exports both, and the [formats section of the homepage](https://www.schemalaw.com/#formats) lists them.

Whichever you choose, the scanned page stays the authority. The layer, the log and the list are all ways of finding your way to it.

## Try it on a real document

Upload a scanned PDF and export it as Word, a searchable PDF or plain text on the free plan. Your documents are never used to train an AI model. [Start free](https://www.schemalaw.com/api/auth/google)