Key content should live in HTML pages, with PDFs offered as an optional download, not the other way round. Google indexes PDFs and AI assistants sometimes cite them, but text extraction from PDFs is less reliable, PDFs carry little context or structure, and they tend to go stale while the HTML gets updated. This post explains how crawlers handle PDFs, when PDFs are fine, the HTML-first pattern, the server headers to use, and how to fix the PDFs you already have.
Can AI crawlers read PDFs?
Search crawlers can read text-based PDFs, but how well a PDF's content survives extraction depends heavily on how it was made. Google lists PDF among the file types it can index, and its crawler documentation says Googlebot fetches the first 64MB of a PDF (versus 2MB for HTML files). So Google can see your PDFs, and AI Overviews and AI Mode can draw on them.
Outside Google, the picture is less documented. AI companies say little about how their search crawlers process PDFs. A few things are true regardless of platform:
- Image-only PDFs have no text. A scanned brochure is a picture of words. Without OCR, there's nothing to extract.
- Layout breaks extraction. Multi-column layouts, tables, sidebars, running headers and footnotes often come out jumbled or interleaved.
- PDFs are single, long documents. AI retrieval works on passages. A 40-page PDF with weak headings gives a system few clean passages to match to a question.
For the general picture of how assistants fetch and use pages, see how each assistant finds sources.
Why PDFs underperform as AI sources
PDFs lose to HTML on almost every signal that helps a page get cited. Even when a PDF is indexed, it has these handicaps:
| Issue | Why it matters for AI answers |
|---|---|
| Weak or missing titles | Many PDFs carry a document title like "Final_v3" or none at all, which is what search results show |
| No site navigation | No internal links, breadcrumbs or context about who published it |
| No structured data | No Organization, Article or Product markup to clarify entities |
| Versions go stale | The HTML page gets updated; the PDF from two years ago stays indexed with old prices or specs |
| Poor passage structure | Headings are often styled text, not tagged headings |
| Weak measurement | Many analytics setups don't track PDF views the way they track pages |
| Dead-end visits | A visitor who clicks a cited PDF lands outside your site with no next step |
The stale-version problem is the one that causes real damage: an old price sheet or spec PDF that keeps getting cited is a common source of outdated information in AI answers.
When PDFs are fine
PDFs are the right format when the document is meant to be printed, signed, archived or used offline. Good uses:
- Legal documents, contracts, terms in their signed form
- Printable forms and checklists
- Datasheets that distributors and procurement teams download
- Academic papers and formal reports, as a companion to an HTML summary
- Brand assets and press kits
The test: if the content answers questions buyers ask an assistant, it needs an HTML home. The PDF can be a convenience copy.
The HTML-first pattern
Publish the full content as an HTML page and link to the PDF from it. For a whitepaper or research report, the HTML page should include:
- An answer-first summary of the key findings in the first few paragraphs
- The findings as text, with real headings and HTML tables for the data
- Methodology, authors and publication date
- A "Download the PDF" link
If the report is gated today, ungate the summary page and keep the form on the full download if you must. See gated content and AI visibility and original research as an AEO asset.
Tell search engines which version is primary
When a PDF duplicates an HTML page, signal that the HTML page is the canonical version. PDFs can't carry a <link rel="canonical"> tag, but Google supports rel="canonical" sent as an HTTP header for non-HTML files.
Apache (mod_headers), one PDF:
<Files "ai-visibility-report-2026.pdf">
Header add Link "<https://www.example.com/research/ai-visibility-report-2026/>; rel=\"canonical\""
</Files>
Nginx, one PDF:
location = /files/ai-visibility-report-2026.pdf {
add_header Link '<https://www.example.com/research/ai-visibility-report-2026/>; rel="canonical"';
}
If you'd rather keep a whole folder of duplicate PDFs out of the index entirely, send X-Robots-Tag: noindex, a method Google has documented for non-HTML files:
location ^~ /files/print/ {
add_header X-Robots-Tag "noindex";
}
Two cautions. First, pick one approach per file, canonical or noindex, not both. Second, don't also block those PDFs in robots.txt: a crawler that can't fetch the file never sees the header. Check the headers with:
curl -sI https://www.example.com/files/ai-visibility-report-2026.pdf | grep -iE "^(link|x-robots-tag):"
More on duplicate handling in canonicals, duplicate content and AI citations.
If a PDF has to stay the primary source
Some PDFs will remain the main version, such as a regulatory filing or a technical standard. Make them as readable as possible:
- Export from the source document with real text, never as a scan; run OCR on legacy scans
- Use tagged PDF export so headings, lists and tables carry structure
- Set the document title and language in the PDF properties (the title is what search results display)
- Use a descriptive filename:
acme-x200-datasheet-2026.pdf, notDS_final2.pdf - Put the publisher, date and version on page one
- Include your site URL and a link to the related HTML page
- Replace updated PDFs at the same URL instead of uploading
-v2copies, and remove outdated ones - Keep file sizes sensible by compressing images
How to find the PDFs you have
Most sites have more PDFs indexed than anyone realizes, often uploaded through a CMS media library years ago. To inventory them:
- Search Google for
site:example.com filetype:pdffor a rough list of indexed PDFs. - Check Search Console's page indexing and performance reports filtered to URLs containing
.pdf. - Search server logs for
.pdfrequests from AI crawlers and search bots. See finding AI crawler activity in your logs. - Ask assistants the questions those PDFs answer and note when a PDF gets cited.
Sort the results into three groups: convert to HTML, keep but canonicalize, and delete or redirect (outdated price lists, old datasheets, superseded policies).
Example: converting a cited whitepaper
This example is illustrative; the details are made up.
A software company finds that Perplexity and Google's AI Overviews cite its 2024 benchmark whitepaper, a 30-page PDF, when people ask about performance in its category. The figures in it are two product versions old.
What they do:
- Publish a new HTML research page with the updated findings, a summary at the top, the data in HTML tables and the methodology.
- Replace the old PDF with an updated PDF at the same URL, adding a canonical header pointing to the HTML page.
- Add the HTML page to the sitemap and link to it from the product and comparison pages.
- Re-run the relevant prompts over the following weeks to see whether citations move to the new page.
The PDF still exists for people who want to download it, but the version assistants find first is the current, readable one.
The bottom line
PDFs aren't invisible to AI search, but they're a weak container for content you want quoted. Put anything buyers ask about into HTML pages with clear structure, keep PDFs as downloadable companions, point duplicate PDFs at their HTML versions with a canonical header or keep them out of the index, and clean up the outdated PDFs that keep feeding old facts into AI answers. Folding this into your regular content audit keeps the problem from coming back.
