You upload a PDF to an AI assistant, ask a simple question, and get an answer that is confidently wrong. The figure it quotes sits in a different row of your table. The summary mixes two unrelated sections. If this sounds familiar, the problem is usually not the AI, and it is not you. It is the PDF itself, and once you understand what is missing from it, the strange answers start making sense.
The symptoms, and what they have in common
The failures follow patterns. Tables come back as walls of run-together text, with values detached from their column headings. Two-column layouts get read straight across the page, splicing sentences from columns that were never meant to meet. Headings lose their rank, so the AI cannot tell a chapter title from a caption. Numbers get attributed to the wrong line.
All of these have one cause: the AI is not reading your document the way you do. It receives the raw text with its positions on the page, and then it guesses. It guesses where a table starts, which column comes first, what is a heading. Where the layout is simple, the guesses hold. Where the document gets interesting, they break.
Why the AI has to guess
Most PDFs only describe what a page looks like: this text at these coordinates, in this font, next to this line. What they do not carry is what the content means. Nothing in the file says “this is a heading” or “these cells belong to the same row.” That information exists in a PDF standard called tagged PDF, a logical structure layer originally created so screen readers could navigate documents for visually impaired users. When it is present, a machine knows exactly what it is reading. When it is absent, every reader that is not human is reduced to inference.
And it is absent most of the time. The majority of PDFs in circulation were exported without tags, because the structure layer was long treated as an accessibility requirement rather than something everyone needs. That calculation has quietly become outdated, for reasons we explore in depth in our analysis published with the PDF Association: the main reader of the world's PDFs is no longer human, and it depends on exactly the layer most documents never included.
What you can actually do about it
Three practical checks, from the simplest case up.
If your PDF is a scan, there is no text in it at all, only a photograph of text; no AI can quote what it cannot read. The first step is optical character recognition, which adds a machine-readable text layer to a scanned document. Note what OCR does and does not do: it recovers the text itself, not the logical structure around it, so it is the floor, not the ceiling.
If your document comes out of Word or another editor, the export settings matter more than most people realise. Modern editors can produce tagged, structured PDFs, but the option is not always on by default. A document exported with its structure intact is a document an AI can follow: real headings, real tables, a declared reading order.
And if your organisation already produces accessible PDFs for compliance reasons, you are further ahead than you think: the structure required by accessibility rules is the same structure AI tools need. We covered what the law now expects from accessible PDFs separately; the practical overlap between that obligation and AI readability is close to total. One effort, two payoffs.
The takeaway
When an AI misreads your PDF, it is rarely being careless. It is being asked to reconstruct meaning that the file never declared. Documents that carry their own structure get read correctly, by assistive technology and by language models alike. Documents that do not, get guessed at. Knowing which kind you are producing, and which kind you are feeding to your tools, is most of the battle.