ChatGPT and other assistants can read a public PDF spec sheet when their web search finds it, and Google indexes PDFs. But a value inside an undated PDF, two clicks from the product page, is harder to find, quote and trust than the same value stated on the page. Put the key facts in text; keep the PDF as proof.
Ask a building materials manufacturer or supplier where the real detail lives and the answer is usually a document. The spec sheet has the tested values. The submittal package has the details an architect reviews. The installation guide has the conditions. The product page has a photo, a paragraph and a download button.
That made sense when the reader was a person who'd click the download. An AI assistant answering a buyer's question has to find the document, read it, and decide which value answers the question.
What we know about AI and PDFs
So the problem isn't that AI can't read PDFs. A PDF makes the job harder in ways a page doesn't.
Google indexes PDFs. It extracts the text from them like any other supported file type, and its AI features draw on that index.
Some assistants retrieve public PDFs during web search and can cite them.
No published study we found measures how often AI answers cite PDFs compared with web pages. Anyone quoting a percentage is guessing.
Where it goes wrong
No date or version. Two spec sheets for the same series, one from 2019 and one current, are both public. An assistant has no way to tell which is right.
The value is in a table image. A scanned table or a chart has no text to extract.
One PDF covers twelve configurations. The U-factor in row four belongs to one size and glazing option. Read without that context, it's wrong for the others.
The PDF lives on a distributor's site. The distributor posted it in 2021 and never updated it.
The link moved. The website was rebuilt and every spec sheet URL changed.
What to do instead
State the important values on the product page, in text: performance class, thermal values, fire rating, the key dimensions. Next to each value, name the standard it was tested to.
Link the document as the proof. The PDF stays, linked from the page, with its title and its date or revision visible.
Put a version or date on every document, and retire the old one instead of leaving both up.
Keep URLs stable. If a document moves, redirect the old address.
Keep the channel's copies in step. If you make the product, tell your distributors and dealers when a document changes. If you sell it, link the manufacturer's current document rather than re-hosting an old copy, and say whose document it is.
Mark up what the page says. Structured data can repeat the values and link the document, but it should never say anything the page doesn't.
Questions
Should we stop publishing PDF spec sheets?
No. Architects and contractors need the document itself for submittals and the project record. The page should carry the key facts, and the PDF proves them.
Does an llms.txt file fix this?
Google says its AI features don't need one. It may help some tools find your pages, but it doesn't fix a missing date or a value in an image.
How do we know which of our PDFs AI is using?
Ask the assistants the questions your buyers ask and look at what they cite. Check My Business does this for your company.