Tag a PDF. Inspect it. Fix the structure.
Upload a born-digital PDF. An untagged, eligible file receives a new visually verified structure; an already-tagged file opens its authored tree without overwriting it. You can inspect either tree, safely correct native paragraph or numbered-heading roles, or explicitly request a separately reviewed visual-AI replacement proposal. Uncertain and complex regions return guidance instead of silently replacing tags. Scans and OCR-derived text are routed without guessing.
Free limits: 5 upload attempts per UTC day per network address, up to 20 pages and 10 MB per PDF. Paragraph and heading roles can be corrected during a private 30-minute browser session. New-tag candidates are rebuilt from the untouched source; existing trees are patched only at the selected roles. Regeneration creates a separate copy only after every rendered page passes reconstruction, independent review, text, number, structure, and drawing checks. Every result remains a review candidate requiring human and assistive-technology checks. Sign up for saved versions, longer documents, alt text, hosted HTML, and the full review workspace.
Public tag workspace
Upload, inspect, correct, download.
- 01Upload
- 02Inspect & edit
- 03Download
Born-digital tag workspace
Untagged PDFs can receive a new reviewed structure. Already-tagged PDFs open their authored tree for inspection, safe corrections, or an optional visual-AI regeneration. Scans and OCR-derived text are routed for specialist review.
- Processed in request memory and never stored by this tool.
- Five upload attempts per UTC day per network address.
- Regeneration uses rendered-page reconstruction plus a separate visual review; uncertain regions are never applied silently.
Read this before you rely on the result
Why a person is still part of accessibility.
What the machine verified
Before a tagged file is offered for download, the drawing operators are compared with the original so the visible pages cannot have changed, a structural audit confirms a usable tag tree with correctly wired content, and the file is re-read through its own tags to prove the text is recoverable at a 90 percent retention gate. These checks are real, and they are also the limit of what a machine can prove.
What only a person can decide
- Whether the reading order matches how a person actually reads the page, not just how text sits in the file.
- Whether a heading level inferred from type size reflects the document's real outline.
- What merged or multi-level table headers mean, and which cells they govern.
- Whether an image is informative or decorative, and what its description should say.
- Whether link text makes sense out of context, and where another language begins.
- Contrast, color-only meaning, and everything a screen reader user actually experiences.
Where this output stands
Eligible simple born-digital files can receive a visually checked tag candidate. Scans, OCR layers, complex content, and uncertainty are refused, and every candidate still needs exact-version human review. This is the industry consensus, not our caveat: the Matterhorn Protocol that defines PDF/UA testing splits its checks into machine-verifiable and human-judgment conditions, the standard checkers pair automation with a required visual check, and the U.S. Department of Justice, extending the ADA Title II deadlines in 2026, observed that generative AI does not yet reliably automate remediation at scale. No tool that claims otherwise is being straight with you, and automated output here never carries a PDF/UA conformance claim.