A scanned document wears a convincing disguise. Open it on screen and the pages look exactly like the paper originals: the letterhead is crisp, the signatures are in place, the tables line up. But underneath that image there may be a hidden layer of recognised text, produced by optical character recognition, and that layer can be wrong in ways you cannot see. A contract clause that reads "shall not" on the page might be stored as "shall not" or, after a bad recognition pass, as "shall riot" or "shall not" with a digit swapped in a date. Search will miss it, copy and paste will spread it, and anyone who relies on the extracted text will be working from a document that is not quite the document they think they have.
That is why checking OCR quality is a separate job from running OCR in the first place. Recognition software reports success when it finishes, not when it is right. A quality check compares what the text layer claims against what the page actually shows, and it does so page by page, because OCR quality is rarely uniform. The first twenty pages of a clean, modern printout may be nearly flawless, while a single faded page near the back, or a page that was scanned slightly crooked, can carry most of the errors in the whole file. The PDF OCR Quality Checker is built for that job: it examines the text layer of a PDF and reports where recognition looks reliable and where it looks shaky, so you can fix the weak pages instead of reprocessing the entire document or, worse, trusting it blindly.
How the tool works
The checker starts from a simple observation: good OCR has recognisable fingerprints, and bad OCR has different ones. Clean recognised text is made mostly of ordinary words, sensible numbers, and normal punctuation, arranged in lines of plausible length. Poor recognition produces the opposite: long strings of characters that do not form words, single letters standing where words should be, digits mixed into the middle of words, and lines that break in odd places. The tool scans the text layer and measures these signals across the document, then summarises them so you can see both the overall picture and the worst pages.
It is worth understanding what the tool is measuring, because the numbers only help if you know what they mean. A high proportion of unrecognised or gibberish tokens on a page usually means the scan itself was poor: low resolution, a skewed page, faint print, or handwriting mixed into typed text. A page where words are mostly fine but numbers look wrong is a different problem, and a more dangerous one, because a wrong digit in an amount or a date changes meaning while still looking plausible. The checker helps you separate these cases by reporting per-page results rather than a single score for the whole file. A single average would hide exactly what you need to find: the three bad pages in an otherwise good two-hundred-page archive.
The tool works on the text layer that is already in the file. That distinction matters. It does not look at the scanned image and re-read it; it audits the text that a previous OCR pass produced. If a PDF has no text layer at all, there is nothing to audit, and the honest answer is that the document needs OCR first, not a quality score. Running the OCR PDF tool on a purely scanned file, and then checking the result, is the correct sequence, and the checker also pairs naturally with the PDF Accessibility Checker, which looks at the wider question of whether a document can be used by assistive technology.
A worked example
Imagine you have inherited a scanned archive of fifty supplier invoices, and you plan to make them searchable so your team can find old orders by supplier name. You run the checker over the file before announcing that the archive is searchable. The overall result looks encouraging: most pages score well, with ordinary words, clean numbers, and consistent line structure. But three pages stand out. Page 12 shows a high rate of non-word tokens and several lines that are only one or two characters long. Page 31 shows repeated digit-letter confusion, where the character "0" and the letter "O" appear to have been swapped in supplier codes. Page 44 is flagged for very short lines and missing spaces, the classic sign of a page scanned at an angle, where the recognition engine chopped words apart.
Each flag points to a different fix, and that is the point of checking before you trust the archive. Page 12, on inspection, turns out to be a carbon copy that was already faint when scanned; rescanning the original at higher resolution is the right move, and no amount of re-running OCR on the same faint image will help much. Page 31 is a clean scan, so re-running OCR with different settings, or simply correcting the supplier codes by hand against the visible page, fixes it quickly. Page 44 needs to be straightened and reprocessed. Without the check, your team would have searched for a supplier, found nothing on those pages, and concluded the invoices did not exist. The archive would have been searchable in name only, with silent gaps exactly where someone was most likely to look.
A second, smaller example shows why numbers deserve special attention. Suppose a page contains the amount "1,250.00" but the text layer reads "1,250.00" with the digit "5" replaced by the letter "S", giving "1,2S0.00". A word-based quality score might barely notice, because one bad token among hundreds is a small percentage. But if anyone searches for that amount, or copies it into a spreadsheet, the error propagates. When you review flagged pages, always check figures, dates, and reference numbers first, even on pages whose overall score looks acceptable. Quality checking is a triage tool: it tells you where to look, and your eyes still do the final verification on anything that matters.
Common mistakes
The most common mistake is treating a single overall score as a verdict. An average of ninety-five percent correct can still hide a page that is fifty percent wrong, and if that page holds the clause or the figure someone needs, the average is meaningless. Always look at the per-page breakdown and investigate the worst pages, not just the headline number.
A second mistake is assuming that a clean-looking page has clean text. Recognition errors are invisible until you select, copy, or search the text, so judging OCR by appearance alone is exactly backwards: the pages that look fine are the ones most likely to be trusted without checking. Make the check a routine step, applied to every document you plan to search, quote, or archive, not just the ones that look rough.
A third mistake is re-running OCR on the same poor scan and expecting a different result. If the image is faint, skewed, or low resolution, the recognition engine is working from bad evidence, and a second pass over the same evidence usually produces the same errors. Fix the scan first: straighten it, raise the resolution if the original is available, improve the contrast, and only then re-recognise. Conversely, do not rescan a whole document when only a few pages are flagged; replace the bad pages and keep the good ones.
A fourth mistake is confusing a quality report with a correction. The checker tells you where the problems are; it does not rewrite the text layer. Flagged pages still need a human decision, whether that is rescanning, reprocessing, or manual correction. Skipping that step and filing the report alongside the document does not make the document any more reliable.
Honest limitations
This tool reports on OCR quality; it does not re-OCR a document, repair its text layer, or restore characters that were never recognised. It cannot see the scanned image the way a human can, so it works from statistical signals in the text rather than a true comparison of image against words, and a passage that is wrong but plausible, such as one real word substituted for another, may not be flagged at all. It also cannot judge whether the original scan is legally sufficient or admissible for any particular purpose; that is a decision for the people responsible for the archive, not for a checking tool. Finally, quality scores are guides, not guarantees. A page that scores well can still contain an error in a critical figure, which is why important documents always deserve a human read of the passages that matter most.
Closing
Trust in a scanned archive should be earned page by page, not assumed from a clean appearance. Check the text layer, investigate the worst pages, fix the scan where the scan is the problem, and verify figures and dates by eye before they travel anywhere important. The PDF OCR Quality Checker gives you the map of where to look; the OCR PDF tool and the PDF Accessibility Checker cover the steps before and after. For the wider question of why scan quality drives recognition results in the first place, OCR Accuracy: Why Scan Quality Decides the Result explains the relationship in detail. A searchable archive is only as good as its weakest page, so find those pages first.