The PDF content stream contains three operators for marking content:

**BMC (Begin Marked Content)** — Opens a marked content sequence with a tag name: ``` /P BMC % begin tagged paragraph content BT /F1 12 Tf 100 700 Td (This is a paragraph.) Tj ET EMC % end marked content ```

**BDC (Begin Marked Content with Properties)** — Opens a marked content sequence with a tag name and a property dictionary (or reference to one): ``` /P <</MCID 0>> BDC % P tag with marked content ID 0 BT /F1 12 Tf 100 700 Td (Paragraph text.) Tj ET EMC ```

**EMC (End Marked Content)** — Closes the most recently opened marked content sequence.

The `/MCID` (Marked Content Identifier) is a non-negative integer that uniquely identifies this marked content sequence within the page. Structure tree leaf elements use the MCID to reference the actual content on the page. The structure element in the tag tree contains: ``` << /Type /MCR % Marked Content Reference /Pg 5 0 R % reference to the page /MCID 0 % the MCID that labels this content on that page >> ```

The full mechanism for linking structure elements to page content works through MCIDs:

1. **Page content stream** — Drawing operators are bracketed by `BDC ... EMC` with a `/MCID N` property. The integer N is assigned by the PDF generator.

2. **Structure element (in the structure tree)** — Contains a `/K` (Kids) array that lists the content it contains. For leaf-level structure elements that directly contain rendered content, `/K` entries are either: - A direct MCID integer (if the content is on the page the structure element is associated with) - A Marked Content Reference dictionary (`/Type /MCR`, `/Pg`, `/MCID`) if the structure element spans multiple pages or needs to explicitly reference a specific page

3. **The link** — When a PDF viewer or accessibility tool encounters a structure element of type `P` (paragraph) with `/K 0`, it knows: "the content for this paragraph is the marked content with MCID 0 on the associated page." It can then render, extract, or announce that content.

This three-layer architecture — structure tree → marked content reference → content stream — is what makes PDF the most structurally sophisticated format for accessible documents. It supports: - Reflow (reordering content by structure tree order, not page position) - Text extraction for assistive technology (following the structure tree order) - Selective content export (extracting just the tagged content, skipping artifacts)

Tagged content problems are among the most common causes of PDF/UA failures. Typical issues:

**Untagged content** — Drawing operators with no surrounding `BDC ... EMC`. The content is rendered visually but is invisible to the structure tree. Screen readers skip it or process it separately in "raw content" mode. Very common in PDFs generated by tools that add content without updating the tag tree (overlays, stamps, watermarks added post-generation).

**Incorrect MCID assignment** — A structure element references an MCID that does not exist on the referenced page, or references an MCID used by a different element. The resulting PDF fails PDF/UA validation with a "content not tagged" or "tag not associated" error.

**Wrong reading order** — MCID values are assigned by the order content is drawn, which may not match the logical reading order. For multi-column layouts, content is often drawn left-column-then-right-column, but the structure tree should reflect reading order, not drawing order. If the structure tree `/K` MCIDs are in the wrong order, screen reader reading order will be wrong.

**Nested marked content without correct nesting** — Marked content sequences must be properly nested. A `BDC ... BDC ... EMC ... EMC` pattern is required if an inner marked content sequence is nested inside an outer one. Mismatched `BDC`/`EMC` operators are a parsing error.

**Appearance stream content not tagged** — Annotations with `/AP` (appearance streams) have their own content streams. Content in appearance streams must also use marked content operators to be accessible; many PDF generators create appearance streams without tagged content.

Tagged content is critical for accurate text extraction — including extraction for AI processing, OCR comparison, search indexing, and accessibility.

When software extracts text from a PDF page, it has two approaches: 1. **Content stream order** — Process drawing operators in the sequence they appear in the content stream. Fast but may produce incorrect reading order for complex layouts. 2. **Structure tree order** — Follow the structure tree, using MCIDs to look up the corresponding content on the page. Produces logically ordered text but requires a properly tagged document.

AI processing pipelines (RAG systems, document Q&A tools) that use content stream order extraction are susceptible to column mixing and reading order errors in complex layouts. Documents with correct tagged content — where the structure tree MCIDs are in logical reading order — produce dramatically better extraction results.

This is why the investment in accessible tagged PDFs benefits AI processing as much as screen reader users: the same structural information that makes a document navigable by a screen reader makes it extractable in the correct logical order by an AI pipeline. See <a href="https://pdftheguide.com/news/pdf-and-ai" target="_blank" rel="noopener">our PDF and AI guide</a> for the full story.