Guide
How to make a PDF readable by AI agents
Make sure every fact in the PDF exists as real text with its label, unit, period and qualifier next to it, and publish a structured text version (Markdown or JSON) of the document at a stable URL that agents can fetch. Agents read text reliably and layouts poorly, so the closer you get to clean, fidelity-checked, addressable text, the fewer errors they make.
How AI agents read a PDF
An AI agent does not see your PDF the way a person does. It extracts the text layer, sometimes runs OCR on images, and then tries to rebuild headings, tables and reading order from character positions. OpenAI’s help center notes that exact values may not be extracted reliably from image-based tables, scanned files or complex layouts, and that chart and image reading inside PDFs is available in specific contexts only; elsewhere only text is read.
An agent that browses the web has an extra problem: it may never open the PDF at all if it can fetch a lighter page, or it may fetch an old copy someone else posted.
Checklist: the PDF itself
- Text layer: export from Word, PowerPoint, Google Docs or your design tool. No scans, no “flatten” or “print to image”.
- Real headings: use heading styles so the structure survives export.
- Simple tables: one header row, units in the header, no merged cells, no tables split across pages.
- Charts: add data labels or a value table; never let a number exist only as a bar height.
- Qualifiers next to numbers: “forecast”, “unaudited”, “pro forma”, “up to” in the same line, not only in a footnote.
- One name per metric throughout the document.
- Dates in an unambiguous form (“March 4, 2026”, not “04/03/26”).
- Reasonable size: OpenAI notes that large or image-heavy files may not be fully analysed.
Checklist: what you publish next to it
- A Markdown version: headings, paragraphs and tables in plain text — the format language models handle most naturally.
- A JSON version if values matter: each metric with its period, currency and qualifier as separate fields.
- A stable URL that does not change when you update the document, so links in emails and decks stay valid.
- A version and status in the file, so an agent can tell a draft from a final.
- An index for the collection, for example an llms.txt file listing your documents with short descriptions (see llmstxt.org).
- Visibility you actually intend: public and indexable, link-only, or restricted.
Conversion is not the same as readability
Running a PDF through a converter gives you text, but not necessarily the right text. Converters can drop a column header, glue two rows together or turn “(1.2)” into “1.2” — and the result looks clean, so nobody notices. For documents where numbers matter, the converted text has to be checked against the source before you publish it.
| Approach | Readable by agents | Numbers checked against source | Stays current |
|---|---|---|---|
| PDF only | Partly | No | Old copies circulate |
| PDF + converted Markdown | Yes | No | Depends on where you host it |
| Fidelity-checked AI Version at a permanent URL | Yes | Yes | Yes, same link |
How The Papers does it
The Papers takes a PDF (text layer with positions; OCR fallback for scanned pages) or a PowerPoint, Excel or Word file and builds one canonical AI Version at a permanent URL: a page for people, Markdown with front matter (canonical URL, version, status, AI Readiness) and JSON with blocks, metrics, qualifiers and page provenance. Your Data Room also gets an llms.txt and an index.
Before anything is published, every number, currency, date and period is checked against the source, and ambiguities go to you to confirm. You can see which AI agents and crawlers read each document and in which format.
FAQ
- Do I still need the PDF?
- Yes, keep it for people who want the designed version. The text version is what agents should read; link to it next to the PDF.
- Is tagged/accessible PDF enough?
- It helps, because structure tags give agents headings and reading order. It does not attach units and qualifiers to values or give you a stable, updatable text address.
- Which format should agents read, Markdown or JSON?
- Markdown for reading and quoting; JSON when an agent needs exact values with their period and currency. Offering both is best.
- Can I keep the document private?
- In The Papers, yes: each document can be public, unlisted (link only, never indexed) or private to Data Room members.