PDF Auditor

BEHIND THE WORKFLOW

How A11yRoo analyzes and tags your PDF

See how document extraction, AI interpretation, structural rules, and output checks work together to create a tagged copy.

  1. 1AssessmentInspect content and identify issues
  2. 2Document structureReview the proposed tags and make optional edits
  3. 3Tagged PDF and final reportCreate a copy and see the remaining findings

WHO HANDLES WHAT

AI interprets. Application rules check. People review.

Extract the document

PDF processing tools

Collect text, images, links, page positions, and existing tags so analysis can refer back to the actual PDF.

Interpret its meaning

AI with document evidence

Propose headings, lists, tables, image descriptions, and reading order using rendered pages, surrounding content, and shared tagging instructions.

Check the tag structure

Application rules and validators

Check tag relationships and source mappings. Apply supported mechanical repairs, such as adding a missing list-body container.

Create and check the copy

PDF writer and output checks

Apply the submitted document structure to a separate PDF, then check content preservation and the resulting tags.

Confirm the experience

Human reviewer

Review meaning and accuracy, and test navigation and reading behavior with assistive technology.

EXPLORE THE THREE STAGES

What happens at each stage

Open a stage for the technical details. The original PDF stays unchanged throughout the workflow.

Assessment

Inspect the PDF and identify accessibility issues, content, and reading order.

  1. Extract evidence and recover text when needed

    PDF.js · pdf-lib · OCRmyPDF · Tesseract

    The original stays unchanged. The processor inventories text, graphics, links, forms, and existing tags. OCR (optical character recognition) can create searchable text for scanned pages in a working copy. Recovered text is checked before it is used; poor source quality can leave gaps.

  2. Interpret pages in document context

    OpenAI multimodal analysis · A11yRoo VACC

    VACC is the app’s visual and contextual analysis process. It relates rendered pages to extracted content and the document-wide plan. Dense columns or tall pages can be analyzed in smaller sections. AI proposes semantic roles—what each piece of content means—and a reading sequence.

  3. Report findings and coverage

    A11yRoo rules · veraPDF · assessment scoring

    Technical checks, contrast findings, and review items form the assessment. The score summarizes the checks and available coverage; it is not an accessibility certification. Incomplete evidence is reported rather than counted as a successful check.

Document structure

Build a proposed tag structure, with optional edits before creating the PDF.

  1. Reconcile the proposed structure

    Document reconciliation · shared semantic contract

    Page proposals are considered together for heading hierarchy, continued lists and tables, and reading order. Every AI request receives the shared tagging requirements. Validators use the same structural definitions to check relationships, while supported automatic repairs correct predictable defects.

  2. Review the document structure

    Editable PDF tags and reading order

    The application prepares editable tags, their relationships, reading order, source references, and alternative text. You can adjust tags or mark an image as decorative before requesting the tagged copy.

  3. Record uncertain decisions without requiring confirmation

    Automatic decisions · semantic review notes

    When interpretations differ, the application selects its best-supported option and records the uncertainty for review. These notes do not require a person’s approval before tagging. Your submitted document structure, including your edits, is the structure used for writing.

Tagged PDF and final report

Write the submitted document structure, check the output, and explain what remains to review.

  1. Check that the document structure can be written safely

    Structural preflight · deterministic repairs

    Preflight means checking the structure and its source references before writing. Deterministic repairs are code-based corrections for known patterns, rather than another AI interpretation. Missing content, unresolved source mappings, or defects that could damage the PDF may still prevent creation.

  2. Write a separate tagged PDF

    pdf-lib · A11yRoo PDF writer

    The writer applies the submitted tags, reading order, alternative text, and associated document structure to a separate copy. Mechanical checks support valid output; they do not require your edits to match the original AI proposal.

  3. Verify the output and deliver the report

    PDF.js · veraPDF · content and structure checks

    Output checks examine whether source content was preserved and whether the written tags reflect the submitted document structure. The final report records findings and review notes. These checks cannot prove that every semantic decision is correct or replace testing the PDF with assistive technology.

RECOVERY AND HUMAN REVIEW

What happens when something is uncertain?

The app preserves completed evidence and uses bounded recovery attempts for unresolved regions. Recovery is not guaranteed: missing or unreliable content and unsafe mappings can leave an assessment incomplete or stop PDF creation. Semantic alternatives are recorded as review notes. A human reviewer still needs to confirm meaning and test the final experience.