Converted PDFs Rank in AI Mode but Refresh Like Images

Google Search now treats a converted PDF as a normal page in ranking and in AI Mode, Martin Splitt said on October 13, 2026. The Search Off the Record episode ran 25 minutes and walked through a converter that sits after rendering, right before a file can enter the index.

Recrawl speed and selectable text now decide whether a terms-of-service file or a hosting whitepaper can show up in an AI answer.

The Converter That Turns a PDF Into a Page

Splitt, who works on Google Search Relations with John Mueller, said Search does not open a PDF in a viewer. A converter in the indexing pipeline looks at the format and writes an HTML version that is as true as possible to the original, because the rest of the web pipeline pretty much only works with HTML.

The converter component is part of what happens in indexing. It happens right after rendering and right before something potentially makes it into the index.

Martin Splitt, Search Relations, Search Off the Record episode 115

That converter is not a PDF special case. Splitt listed converters for rich text files, document files, Excel sheets, and Lotus Notes documents for some reason. Images and video still go down other pipes. Once the HTML version exists, links inside the PDF are ordinary links, which is why Mueller said Google can follow them the way it follows links on a web page.

The recording is on the Google Search Central YouTube channel, and the PDFs in Search episode page carries the same chapter list, including rasterized pages, binary copies, and AI Overviews. The official episode transcript is the clean record of the wording.

Splitt was less sure about pictures sitting inside the file. He thinks Google may see that an image is there, but the usual Image Search path wants a URL, and an embedded PDF image does not have one. He said that may change. Until it does, charts in a whitepaper should not be counted on for image results.

The 64MB Fetch Cap the Hosts Could Not Name

Mueller asked how size limits work for PDFs. Splitt said fetching a PDF probably has a different limit than a web page, then admitted he did not know the specific limit for PDF files. He also said a web-page-style limit may apply to the HTML version after conversion, and he was not sure where those cuts are made.

Google already printed the fetch numbers. The Googlebot help page, last updated February 3, 2026, says the crawler takes the first 2MB of a supported file type and the first 64MB of a PDF. The cap is applied on uncompressed data. Once the cutoff is hit, Googlebot stops and sends only the downloaded part for indexing.

Gary Illyes restated the same pair on March 31, 2026 in a Search Central post on crawling bytes. Googlebot fetches up to 2MB for any individual URL excluding PDFs, including the HTTP header, and the PDF limit is 64MB. Bytes after the cutoff are not fetched, not rendered, and not indexed.

THE PDF PIPELINE DATES

1. February 3, 2026: Googlebot help states the 2MB cap for supported files and the 64MB cap for PDFs.
2. March 31, 2026: Illyes’s crawler post repeats those figures and adds a 15MB default for crawlers that set no limit of their own.
3. October 13, 2026: Splitt and Mueller describe the converter on episode 115 and say they do not know the specific PDF size limit.

Splitt’s caution still holds in one sense. He said the limit can be changed, and that is why Google is not keen on a figure that will float around for a long time. The help page is still the number in writing today, and it is much larger than the HTML fetch, which is why a scanned manual can get in while a bloated HTML page gets cut.

FETCH CAPS FOR SEARCH

For most hosting blogs the 2MB HTML cap never comes up. It does come up for a knowledge-base export, a legal pack, or a product manual that still lives as one giant PDF. The first 64MB is a lot of text. It is not the whole file if someone bundled high-resolution scans behind the essay.

Treat a PDF Like an Image When You Update It

Ranking was the easy part of the recording. Splitt said indexing and ranking is pretty much the same as any other document. Mueller then described how often Google comes back, and that is where a PDF stops behaving like a page.

So our systems are not going to refetch the PDF file every couple of minutes, because they basically assume that this file is probably not going to change for a really long time.

John Mueller, Search Relations, Search Off the Record episode 115

He compared that refetch habit to images. The systems also do not want to pull lots of large files off a server. So a terms page you edit in HTML can move in days, while the same words saved over a PDF at a stable URL can sit in the old form in Search, and in AI Mode, until a rare recrawl.

That is the cost that the “PDFs can rank” line does not carry. If your acceptable-use policy, SLA, or price sheet is the file customers actually download, an in-place overwrite is a long-lived copy in Google’s index. A new filename, or an HTML page that holds the live text, is the way the index sees the new words.

People who had watched PDFs drop from results in the weeks before the episode did not get an explanation on the tape. The hosts were asked, in public replies to the recording, whether they had covered that drop. They had not. The converter talk does not tell you why a file that used to rank went missing, only how a file that is still in the index is stored.

Scanned Pages Leave No Text to Index

Splitt called rasterized pages the worst problem. A PDF can save each page as an image, and then the converter has no words to turn into HTML. Mueller’s practical test is the same one you can run without Search Console: select the text and paste it into a text editor. If the words appear, Google can use them. If you get a blank page, so does the converter.

That failure is common on hosting sites that scan signed contracts, stamped invoices, or “print to PDF” menus of plan comparison tables. The file looks fine to a customer. It ranks for nothing because there is no text layer. Splitt also tied the same defect to accessibility: screen readers cannot read a picture of a paragraph unless the generator went out of its way to include real text.

Password-protected files stay closed to crawlers, which is the right setup for customer exports and anything that should never be public. The hosts also warned against saving an llms.txt file as a PDF, even if you like llms.txt. A whole book in one PDF, Mueller noted, has the same problem as a whole book on one HTML page: the unit is too large for the way people search.

HOW A PDF FAILS THE CONVERTER

• Rasterized pages: Each page is an image, so highlight-and-copy returns nothing and ranking text never appears.
• Password lock: Crawlers cannot read the file, so private invoices and client folders stay out of the index.
• Silent overwrite: The same URL is treated like a static image, so an edited policy can remain stale for a long time.
• Past the 64MB mark: Bytes after the fetch cutoff are ignored, including text you stacked behind heavy scans.
• llms.txt as a PDF: The hosts flagged that packaging as a bad idea for AI crawlers that want plain text.

If the content can live as HTML, Mueller’s bias is to put it there. If the content has to be a PDF, the text layer is the ranking work. Design polish does not substitute for words the converter can lift.

Why Two Copies of the Same PDF Confuse Google

Mueller said PDFs are a binary blob compared with HTML pages. If you publish that blob and another site downloads it and hosts the same bytes, search engines see one-to-one identical files. Canonicalization is harder than it is with two similar HTML pages, because there is no “almost the same article” to reason about. It is the same object.

That hits research PDFs, vendor whitepapers, and the manuals hosting companies leave in a public /docs folder. Mirrors are easy. The episode’s own notes point to a rel=canonical HTTP header, or an embedded header link, so Google can tell which host is the original. Without that, the copy that ranks may not be yours.

Links still pass through the conversion, so a PDF is not a dead end in the link graph. It is also a duplicate trap if the same file is the downloadable twin of an HTML article. When a company publishes both, the HTML page is the one that recrawls, accepts structured data, and works on a phone. The PDF is the print artifact. Treating them as two equal “pages” is how you end up arguing with a blob.

PDFs Inside AI Mode With No Extra Markup

Mueller asked whether converted PDFs can rank normally. Splitt said yes, indexing and ranking is pretty much the same as any other document. Mueller followed up on AI Overviews, AI Mode, and the rest. Splitt’s answer was that internally it is a normal HTML page, nothing wild. He compared it to JavaScript: Google does not track whether something was generated by JavaScript, and it does not care if something was a PDF originally.

Mueller then took the point past Google. He said they had already talked about GEO, and that at Google, PDFs do work for it. He has seen them referenced in other AI systems and assumes most of those systems want to train on the content and will use it for grounding too. There is no extra PDF-only GEO switch. If the converter can read the file, the same systems that cite HTML can cite the converted page.

That is useful for a hosting knowledge base that still ships PDF runbooks, or for a registrar that publishes a policy pack as a download. It is not a reason to hide the live text in a file Google refetchs like an image. AI Mode can quote a stale PDF as confidently as a fresh one.

HTML Still Makes the Same Content Easier to Use

Splitt’s publishing rule was blunt. If the content warrants a PDF because you need print layout, a form, or a menu the printer already made, upload it. A holiday lodge registration form or a restaurant menu is better as a PDF than as nothing. He still prefers HTML as a flexible format.

Mueller put the same choice in user terms. If the option is a menu as a PDF or no menu at all, he would rather have the menu. If the option is a PDF versus a normal web page, he likes the web page because it is easier to interact with. He also said that if you can put the same thing in HTML as in your PDFs, that makes everything a little bit easier.

For a hosting site the split is practical. Keep the article, changelog, and pricing table in HTML so recrawl, phones, and AI answers see the live text. Keep the PDF for the printable guide, and give it selectable text, working links, and a new URL when the words change.

Frequently Asked Questions

Does Google Index PDF Files in Search?

Yes. Google lists Adobe PDF among its encoded types including Adobe PDF, and it usually reads the type from the Content-Type HTTP header, falling back to the file extension if that header is missing or wrong. The same encoded list includes Word, Excel, PowerPoint, RTF, and OpenOffice files, which is why Splitt’s converter talk was never only about PDFs.

What Is the Googlebot Size Limit for PDF Files?

The published Search fetch is 64MB for a PDF and 2MB for other supported files. Crawlers that set no limit of their own default to 15MB, which is a separate platform default, not a second PDF cap. Splitt said the number can move, which is why the hosts would rather not freeze a figure in people’s heads even though the help page already states one.

Can a PDF Appear in Google AI Mode?

Yes. Splitt said the converted file is internally a normal HTML page, and Google does not track whether it started as a PDF, the same way it does not track whether HTML came from JavaScript. Mueller added that he has seen PDFs referenced in other AI systems and assumes those systems want the files for training and for grounding, not only for classic blue links.

Can Google Index a Password-Protected PDF?

No. A password keeps crawlers out, which is the correct setup for invoices, customer exports, and any file that should stay off the public index. If a public PDF is already in results and you want it gone, Google’s help points to a noindex rule on a URL it can fetch; blocking the file in robots.txt can still leave the URL in results without its text.

Do Links Inside a PDF Count Like Web Links?

Yes, because they survive in the HTML conversion and the rest of the system treats that page as HTML. For copies of the same file on other sites, the episode points to a rel=canonical HTTP header or an embedded header link so Google can tell which host is the original, which HTML pages get for free from ordinary canonical tags.