OCR PDF check: compare scanned fields with copied text

What are you checking before the spreadsheet?

An operations assistant receives a scanned order sheet. It looks readable, so copying its contents into a spreadsheet seems straightforward. But the text selected from the page may contain different characters from those visible in the scan. A neat spreadsheet can therefore start with an incorrect identifier or amount.

OCR means optical character recognition: software turns pictures of writing into usable text. The text layer is the selectable writing stored in the PDF. This exercise checks that layer against the page image before you calculate anything or ask an assistant to interpret it. It does not evaluate an AI answer.

Choose one page and three fields that matter to your next action. An item reference, a date and a quantity are enough. Your deliverable is a saved comparison record identifying matches, differences and anything you cannot read. The result covers those fields, not every page in the file.

For the wider workflow, start with the practical artificial intelligence guide. A more capable assistant does not remove your responsibility to check the input you provide.

Can you copy the text already?

Keep the received file unchanged and make a working copy with a recognisable name. Open the copy and try selecting one line. On Windows, search Start for Notepad, open the application and create a new blank document. Paste the copied line there, not into a spreadsheet that might automatically change dates or remove leading zeros. Record what actually appears.

There are several possible observations: nothing selects, an entire image selects, or readable text copies. Do not interpret every selection problem as proof that OCR is missing. File permissions or your viewer may need investigation. Ask the owner for an appropriate source instead of trying to bypass restrictions.

If the text copies, move straight to comparison. If it does not, check whether an existing application you are authorised to use includes OCR. Adobe documents an Acrobat route through Edit for scanned PDFs and advises reviewing recognition output. Features depend on your available access; this exercise does not require buying a subscription.

Keep conversion and editing separate. You are obtaining a representation of the source, not changing the source order. If an amount looks surprising but is clearly printed, preserve it and raise a question. A plausible business value is not evidence of what the page says.

Build a three-field evidence record

Write the location before writing the value. Use the filename, viewer page number and a useful row label. If the page has a different printed number, include both. This lets another person locate the same field without guessing whether you counted a cover sheet.

Read the visible field first, then inspect the pasted text. Keep both versions in separate columns even when they agree. The following is a deliberately constructed training example, not a reported software test, customer document or measured error rate.

The identifier needs a correction to the extraction if the visible source is unambiguous. The date matches in this example. The quantity requires a more careful decision: punctuation can express different number conventions. Do not convert it automatically before confirming how the document uses separators.

Add the unit to your record. A quantity without its unit may be copied accurately and still be unusable. If a heading applies to several rows, retain that heading in the note. You should not have to reopen the file merely to remember whether the field meant packs, individual items or hours.

What should happen when a field is unclear?

Use three statuses: matched, corrected extraction and unresolved. A corrected extraction means that the copied text was changed to reflect a clearly readable source. It does not mean you corrected a supplier's document or approved its commercial contents.

When the image is blurred, write unresolved and request a clearer original. Do not use neighbouring entries to invent the missing character. An identifier that normally follows a pattern can still contain an exception. Your record should distinguish what you observed from what you suspect.

W3C's PDF7 technique describes checking converted text, including its completeness and reading order. Our three-field exercise is intentionally narrower than a full document or accessibility assessment. Once you move beyond this sample, include headings and row order so values remain associated with the correct labels.

Keep a separate next-action column. A matched field may be ready for the next task. An unresolved reference may require a replacement scan. An extraction correction may require the downstream spreadsheet to be updated too. Give each follow-up an owner rather than treating every problem as a reason to repeat OCR.

If the next step is a purchasing comparison, use the source-first supplier quote worksheet only after the relevant fields and units have been checked.

Save a note someone else can reproduce

The site's published PDF review guide recommends checking answers against their source and separating checked, corrected and unresolved items. This exercise applies that documented working method to the PDF input itself. It is not a claim that Anar performed this OCR test or obtained a particular client result.

Read the PDF analysis guide for the later interpretation stage. For now, copy this note into a text file:

Source filename and version:
Viewer page / printed page / row label:
Field name and unit:
Visible source value:
Copied text:
Status: matched / corrected extraction / unresolved
Checked replacement text, if any:
Reason and next action:
Fields and pages not checked:
Reviewer and date:

Press Ctrl+S, choose Documents and save as ocr-check-v1.txt. Close the editor. In File Explorer, open Documents and double-click the saved file. Confirm that the source filename and the comparison are still understandable. A note kept only in a temporary chat window is easy to separate from the file it describes.

Avoid a vague summary such as “PDF checked”. Write “page one, three selected fields checked; remaining rows not reviewed”. This gives a colleague a useful boundary. When a clearer scan arrives, create a new note naming that version and repeat the affected comparisons rather than silently replacing the earlier evidence.

The document version approval guide helps keep references tied to the exact copy. Version approval and accurate OCR are separate checks, so keep both when the workflow needs them.

Finish a small useful check in fifteen minutes

Spend the opening minutes choosing the page, preserving the original and naming the three fields. Next, test copying or use an available OCR function. Compare the fields, record the outcome, then save and reopen your note. This is a suggested timebox, not a measured performance promise.

If OCR access or a readable source is missing, your useful output is a precise request: identify the page, the field and what prevents the comparison. Do not spend the entire time trying unrelated converters. Asking for the original text document may be the simplest next step.

Before handing over, inspect any correction where leading zeros, decimal separators or units matter. Paste the checked replacement into plain text once more. Confirm that the downstream destination has not changed its format. A spreadsheet displaying 001 as 1 creates a different problem from an OCR error.

You can use AI to draft a list of questions about the extracted text, but not to certify a page it has not seen. Ask it to flag uncertain fields and avoid filling missing values. Keep its suggestions outside your verified replacement column until you inspect the source.

The finished task is modest and concrete: one saved record, three traceable comparisons and a next action for unresolved entries. That is enough to prevent a small input problem from becoming an unexplained difference later.

Before sending the file, perform the separate attachment or shared-link access check. A correct text layer does not prove that the intended recipient can open the right copy.