How to Optimise PDFs and Unstructured Assets for AI Citation

Most companies keep their most quotable material in the least quotable format. The original research, the benchmark numbers, the annual figures: all sitting in a PDF that a language model either skips or reads badly. Meanwhile a competitor''s thinner blog post gets cited because it is plain HTML with clear headings.

Can AI engines read and cite PDFs?

Yes, but unevenly. A text-based PDF with a real heading structure can be parsed and quoted. A scanned document, a design-heavy report with text in images, or a multi-column layout often extracts as scrambled text or not at all. The format is not the barrier. The construction is.

Why PDFs fail at parsing

• Text stored as images. Scanned pages and exported design files carry no selectable text, so there is nothing to extract without OCR.
• Multi-column layouts. Extraction often reads straight across the page, interleaving two columns into nonsense.
• No structural tags. Visual headings made by enlarging text carry no heading role, so the document arrives as one undifferentiated block.
• Numbers trapped in charts. The figure a model would most like to quote exists only as pixels.
• No stable URL. Gated or session-scoped files cannot be cited even when parsed.

The fixes that matter most

Publish an HTML version alongside the PDF

This single change outperforms every other tactic. Put the full content on a normal web page with proper headings, and offer the PDF as the formatted download. The HTML version is what gets parsed, quoted and linked. The PDF remains the thing people email internally.

Tag the document properly

Use real heading styles rather than manually enlarged text, add alternative text to figures, and set document title and language in the file metadata. Tagged PDFs extract in reading order, which also makes them accessible to screen readers.

Put every key number in text

Any statistic worth citing should appear in a sentence, not only inside a chart. Add a short data table or a plain-language summary line under each figure. If the number cannot be selected with a cursor, assume it cannot be quoted.

Lead with a summary block

Open with a hundred-word summary of the findings, in plain sentences, near the top of both versions. Retrieval systems favour self-contained passages, and this is the passage most likely to be lifted into an answer.

Make it citable

Give the asset a permanent URL, a publication date, a named author and a clear organisation attribution. Ungated is not always commercially possible, but a gated file is functionally invisible to a model, so consider ungating the findings and gating the deeper appendix or template.

The trade-off with lead generation

Gating exists for a reason and this advice complicates it. The workable compromise is to publish the findings openly, in HTML, with the data and the summary, and reserve the gated download for the parts a buyer wants for their own work: the raw dataset, the spreadsheet model, the implementation template. You keep the lead capture and you stop being invisible to every assistant your buyers now ask first.

Our guide to getting cited by AI engines covers the same principles for ordinary pages, and a free audit will show which of your existing assets engines can actually read.