Short answer. Partly, and not in the way most teams assume. A born-digital PDF is readable because it carries a text layer; a scanned one is a picture of a document and yields nothing without OCR. Images are mostly understood through the text around them — filename, alt text, caption, nearby copy. Video is read almost entirely through its transcript, title and description, which is why transcripts have emerged as one of the strongest correlates of AI visibility. Every format earns citations through text.
Key takeaways
- A born-digital PDF carries an embedded text layer and can be parsed directly; a scanned PDF is a sequence of images and contains no machine-readable text without OCR.
- Answer engines mostly understand images through surrounding text — filename, alt attribute, caption and nearby paragraph — not through vision alone.
- Video is read almost entirely through its transcript, title and description; a video with no corrected transcript contributes very little to citation.
- Auto-generated captions degrade sharply in lower-resource languages and regional accents, so transcript coverage is language-specific, not universal.
- PDF reading order and table structure frequently break during text extraction, which is why an HTML version of a document is more citable than the PDF itself.
What can each format actually contribute?
All three formats can contribute, but only through text. Every format that gets cited does so because something in it was readable as text.
That single principle explains most of the confusion here. Answer engines synthesise text answers, so a format earns a citation when it yields extractable, attributable text. The differences between formats are really differences in how much text they expose, and how reliably. A text layer is the machine-readable string of characters embedded in a file, as opposed to the visual appearance of those characters on a page.
There is a commercial reason to care. Analysis of AI Overview inclusion has found that multimodal content — text combined with images, video and structured data — correlates strongly with inclusion, an early-mover advantage as answer engines expand into processing non-text content. Multiformat pages that expose that content as text tend to perform well; formats that hide their content do not. For a broader rubric on what earns a citation across formats, see what actually gets you cited by AI answer engines.
Why do PDFs get read but rarely cited well?
Because a PDF is a print format pretending to be a web page. The text is often there; the structure an engine needs to quote it accurately frequently is not.
Start with the split that decides everything. A born-digital PDF, exported from a word processor or design tool, contains an embedded text layer and can be parsed directly. A scanned PDF is a sequence of images and, without OCR, contains no machine-readable text at all. Every whitepaper, report and datasheet on your site falls into one of those two categories, and many organisations do not know which.
Even when the text is present, three structural problems reduce citability.
| Problem | What breaks |
|---|---|
| Reading order | Multi-column layouts, sidebars and pull quotes parse in the wrong sequence, so extracted passages come out scrambled and unquotable |
| Tables | A visual grid flattens into a run of numbers with no relationships preserved — a particular loss, because tabular data is otherwise highly quotable |
| Missing context | No reliable heading hierarchy, no schema markup, no publish date in a machine-readable field, no internal links. The page hosting the file may have all of that; the file does not |
Document-parsing vendors building retrieval pipelines report exactly this: plain OCR can extract raw text from a scan but produces poor reading order and broken tables, which damages chunking and answer quality even when the raw text is technically present. This is one reason technical AEO structure treats HTML as the primary surface and the PDF as a secondary artefact.
The practical conclusion is not to abandon PDFs. It is to stop treating them as the primary version of anything you want cited. Publish an HTML page carrying the same content, properly structured, with question-shaped headings as described in question headings and answer-first writing, and offer the PDF as the download. The HTML earns the citation; the PDF serves the reader who wants to print it.
How do engines actually understand images and video?
Through the text attached to them. Vision capability exists, but the reliable path to citation runs through captions, alt text, transcripts and descriptions.
Images. Modern models can describe an image, but an answer engine deciding whether to cite a page is working mostly from the text around it: filename, alt attribute, caption, nearby paragraph, structured data. This is why charts perform better when their key finding also appears in the caption or body copy. An image carrying a statistic that no sentence on the page repeats is a statistic the engine cannot quote, and images that do surface in AI responses are typically linked back to a source, so descriptive attribution matters.
Video. The evidence here is the strongest in this whole area. YouTube consistently ranks among the most-cited domains across answer engines, and research across 75,000 brands found that brand mentions in YouTube video titles and transcripts correlated more strongly with AI visibility than any other signal studied. Practitioner analysis points the same way: transcript density, chapter markers and structured descriptions all lift extractability.
Read that carefully, because it is easy to misread. It is not that video content is favoured. It is that a video with a rich transcript is a text document with a video attached, and the transcript is what gets read. A video with an auto-generated, uncorrected or absent transcript contributes very little.
The multilingual dimension is where this gets commercially interesting, and it connects to the pattern described in the English bias in AI search: a transcript exists in one language unless someone produces the others, and automatic captioning degrades sharply in lower-resource languages and regional accents. A company with excellent English transcripts and nothing else is invisible the moment a customer asks in Bahasa Indonesia or Arabic. Producing accurate transcripts and captions across languages, as described in a multilingual content pipeline AI engines cite, is work Lifewood does through native speakers across its 40+ delivery centres in 30+ countries — in this context a visibility investment rather than an accessibility afterthought.
What should you fix first?
Fix these in order, because the effort-to-effect ratio differs sharply.
- Audit which PDFs are scanned. Open each one and try to select text. If you cannot, no engine can read it either. Run OCR — or better, republish as HTML.
- Give every important PDF an HTML equivalent. Same content, question-shaped headings, an answer near the top of each section, real HTML tables rather than images of tables.
- Write alt text that states the finding, not the file. "Bar chart showing rural internet use at 58% against 85% urban" is useful. "chart1.png" is not.
- Repeat the key number in prose. A statistic that exists only inside an image effectively does not exist for citation purposes.
- Publish corrected transcripts, not auto-captions. Add chapter markers and a structured description. Treat the transcript as a page in its own right, following the same crawler-access checks covered in which AI crawlers to allow and which to block.
- Do the same in every language you sell in. Transcripts, captions and alt text are language-specific assets; coverage in one language buys nothing in another.
- Check crawler access to your media directories. A
robots.txtrule blocking/assets/or/downloads/quietly excludes everything above, and nothing in your analytics will report it.
A managed AEO programme typically runs this audit as its first pass; see AEO/GEO providers for how that work is scoped.