Skip to the indexed answers
Digital Marketing Index

usdigitalmarketingindex.com SEO

How can you tell whether text in a scanned PDF is accessible to search engines?

By trying to select and copy the text, then checking whether what comes out is the text on the page. A scanned PDF is a photograph of a document. Unless a text layer has been added, it contains no characters at all, only pixels that look like characters to a person. If a word cannot be selected, there is no text to read. If it can be selected, a text layer exists, and the next question is whether that layer is correct, since the layer produced by optical character recognition may contain errors that are invisible on screen. Google described its own rule of thumb in a 2011 post: “if you can copy and paste the text from a PDF document into a standard text document, we should be able to index that text” (Google’s 2011 post on PDFs in search results). The copy-and-paste test, followed by a comparison, answers the question for any file.

Image-only versus searchable text

Three kinds of PDF look alike on screen.

Type How it was made Contains characters?
Born-digital Exported from a word processor or design tool Yes, as real text
Scanned, image-only A scanner or camera produced page images No
Scanned with a text layer A scan processed with OCR, which placed recognized text invisibly behind the image Yes, as recognized by the software

A comparison of the first and second, as a reader would test them:

Test Image-only document Selectable-text document
Drag the cursor across a line A box is drawn over an image area, or nothing is selected Individual words highlight
Select all and copy; paste into a text editor Nothing, or an empty paste The document's text appears
Search within the viewer for a word visible on the page No results The word is found and highlighted
Zoom in Characters become blurred or pixelated Characters stay sharp, if born-digital; a scan with a text layer still shows a blurred image
File size relative to page count Large Smaller, if born-digital
Document properties, fonts No fonts listed Fonts listed

The third row is the quickest check. If the viewer’s search cannot find a word you can see, the file has no usable text for that word.

In the same 2011 post, Google said that where text is embedded as images, “we may process the images with OCR algorithms to extract the text”. The word is “may”. That is a statement from 2011 about a possibility, and it is no basis for relying on a search engine to read an image-only file. A publisher who wants the text read should supply the text.

Text layer and OCR limits

A text layer makes text selectable. It does not make it right. What the viewer displays is the scanned image. What is selected, searched, and extracted is the hidden layer. The two can differ without any visible sign.

OCR error Effect
Character confusion: l and 1, O and 0, rn and m Wrong words and wrong numbers
Missed or merged words Text runs together or drops out
Wrong reading order Columns interleaved; headers mixed into body text
Tables Cells read in the wrong sequence; figures detached from labels
Hyphenation and line breaks Words split
Accents and non-Latin scripts Wrong characters if the language was not set
Handwriting, stamps, low-contrast or skewed scans Garbage or nothing
Headers, footers, page numbers Repeated through the text
Small print and footnotes Misread
Formulas and symbols Lost

So “selectable” is the first test and not the last. A document whose text layer reads “the pemnit fee is S1O0” is technically searchable and practically wrong: a search for “permit” will not find it, and a reader using a screen reader will hear nonsense.

Accessibility is a separate property again. A text layer gives assistive technology characters to read. A properly accessible PDF also has tags that mark headings, lists, tables, reading order, and alternative text for images. A file can pass the copy test and still be unusable with a screen reader.

Property Means
Selectable text A text layer exists
Accurate text The layer matches the page
Correct reading order The text comes out in the order a person reads
Tagged structure Headings, lists, and tables are marked for assistive technology
Document language and title set Declared in the file

Testing copied text

A verification procedure:

  1. Open the PDF in a standard viewer.
  2. Search for a distinctive word visible on the first page. Note whether it is found.
  3. Select all the text and copy it. Paste into a plain text editor.
  4. If nothing pastes, the file is image-only. Stop: it needs OCR or replacement.
  5. If text pastes, compare it with the pages.

What to compare:

Check Method Looking for
Completeness Compare the length of the pasted text with what the pages contain Missing pages or sections
Accuracy Read a paragraph from the start, the middle, and the end against the image Misrecognized words
Numbers Check every figure in a sample table against the image Digit errors
Names and technical terms Search the pasted text for key terms known to be in the document Terms not found because they were misread
Reading order Read the pasted text straight through Columns or captions out of sequence
Tables Compare a table as pasted with the original Scrambled cells
Every page Search for a word from each page Pages with no text layer
Language Look at accented and special characters Substitutions

Mixed documents exist. A file can have real text on some pages and images on others, for instance a typed report with scanned appendices. Test more than the first page.

Further checks:

Check Method
Text extraction with a tool Run a command-line text extractor on the file and read the output; this is close to what an automated system receives
Accessibility checker Use the checker in a PDF editor to report missing tags, reading order, and language
Screen reader Listen to a page
Security settings Check document properties for password protection or restrictions on content copying; the 2011 post says Google can index text from PDFs "provided they're not password protected or encrypted"
Search engine check After the file has been public for some time, search for an exact sentence from it in quotation marks, restricted to the site; presence confirms the text was read, and absence does not prove it was not

Remediation and QA

Finding Remedy
Born-digital source exists Export a new PDF from the source document; do not scan a printout
Image-only scan, source lost Run OCR; then proofread the text layer
Text layer with errors Correct the layer in a PDF editor, or re-run OCR at higher quality with the right language, then proofread
Poor scan Rescan at higher resolution, straight, with good contrast
Important content Publish it as an HTML page as well, with the PDF as a download
No tags Add tags, reading order, language, and title in a PDF editor
Tables Rebuild as real tables, or provide the data in HTML
Restricted or encrypted Remove restrictions that block text extraction, if the document is meant to be public

For anything that matters, the durable answer is HTML. A web page with the document’s content is readable by every search engine, usable on a phone, and accessible when built properly. The scan can stay as the authentic image of the original.

QA after remediation:

Check Pass condition
Copy test on every page Text pastes for all pages
Proofread Sample from each section matches the image; all numbers in key tables verified
Search test Key terms found within the viewer
Extraction tool Output readable, in order
Accessibility checker No critical failures
Document properties Title, language set; no copy restriction
File served correctly 200, PDF content type, not blocked by robots.txt
Record Who proofread, when, and what was corrected

Proofreading has to be done by a person for documents where accuracy matters, such as legal texts, technical specifications, and anything with figures. OCR accuracy varies with the scan, and a small error rate over a long document is still many errors.

Scale the effort. For an archive of thousands of scans, test a sample to learn which batches are image-only, run OCR on those, and proofread the documents people use.

Once the PDF’s text is sound, a further question is whether it should be found in search in place of, or alongside, an HTML page with the same content, covered in the guide to how a PDF and its equivalent HTML page should avoid competing for the same search intent.

The test: select the text, paste it somewhere plain, and read it against the page; if nothing pastes there is no text, and if what pastes is wrong the text is not the document’s.

Link trail

Reader comments on this answer

Add your comment to this answer

Your email address will not be published. Required fields are marked *