seo-headings-structure (M7c)

A clean, single-rooted heading tree and proper semantic landmarks make a page's outline machine-readable — the same structure that AI engines use to extract answer blocks and that search engines use to understand content scope. Schema/entity context: references/schema-tier1.md.

Inputs

Work from the PageSnapshot named in your dispatch envelope: read parsed from <run_dir>/pages/<slug>.json (headings[] with region, landmarks); Grep pages/<slug>.html for verbatim evidence; site artifacts live in <run_dir>/site/{robots.json,sitemaps.json,discovery.json}. Deterministic findings already emitted by audit.mjs are listed in <run_dir>/findings.deterministic.json — do not re-emit those ids; add model-judged findings only. If invoked directly with a URL/path and no snapshot exists, first run node "${CLAUDE_PLUGIN_ROOT}/scripts/snapshot.mjs" <target> --out "${CLAUDE_PLUGIN_DATA}/runs" and use the printed snapshot path.

Audits

Working from the PageSnapshot (parsed_rendered when render.used is not none, else parsed):

1. Single H1: exactly one <h1> per page. Zero H1s, or multiple H1s, are both findings.
2. No skipped levels: nesting descends one level at a time (an <h2> may be followed by <h3>, not directly by <h4>). Build the outline and flag any jump.
3. Descriptive text: each heading names its section in human terms — flag empty headings, headings used purely for visual styling, and generic placeholders ("Section 1", "Welcome").
4. Semantic landmarks: detect <main>, <article>, <section>, <nav>, <header>, <footer>; flag when headings live inside non-semantic <div> soup or when more than one <main> exists.
5. Outline ↔ landmark agreement: sectioning elements should carry their own heading; orphan landmarks and headings outside any landmark are findings.

Good heading structure feeds AI answer-block extraction: clear H2/H3 boundaries map to candidate quotable passages.

Fixes

• Heading-level restructuring — re-leveling H2-H6 to remove skips or to nest correctly (fixable: proposed). This alters the visual hierarchy, so it requires per-item accept; emit a fix_preview diff, never an auto-write.
• Add a single missing H1 — when the page has no H1, propose one sourced from the populated <title>/og:title or the obvious page-title element (fixable: proposed, per-item accept — adding a visible heading is editorial). Never fabricate H1 text from thin content; ask the user if the source is ambiguous.
• Wrap in semantic landmarks — recommending <main>/<article> adoption is advisory (structural refactor the tool does not write).

Mark each finding's fixable field consistent with schema/finding.schema.json (auto / proposed / advisory). Never fabricate a heading's text or invent an H1 from thin content — ask the user or leave a clearly-marked TODO placeholder.

Verification

• Method: dom_assert. Re-parse the snapshot, rebuild the outline, and assert the specific condition (one H1, no level skip, landmark present).
• Offline reproduce: node "${CLAUDE_PLUGIN_ROOT}/scripts/parse-html.mjs" --snapshot <pages/<slug>.json> (or --url <u> / --file <path>) — returns the headings block (counts per level, h1_count, and skipped_levels), which counts H1s and walks the H1-H6 tree for skipped levels.
• When the rendered DOM tier is required but unavailable (heading injected client-side and only raw_html_path is present), status is needs_api, never a false pass.

Findings

Emit findings per schema/finding.schema.json. Examples:

• M7c.h1.multiple — more than one <h1> on the page (status fail, severity 3, fixable: proposed, axis search, confidence established). evidence.observed quotes each H1's text and selector; verification.reproduce is the parse-html.mjs command above (inspect h1_count).
• M7c.outline.skipped_level — an <h2> followed directly by an <h4> (status fail, severity 3, fixable: proposed, axis search, confidence directional). evidence.observed quotes the two adjacent headings and their levels.
• M7c.landmark.no_main — no <main> element; primary content sits in <div>s (status warn, severity 3, fixable: advisory, axis search, confidence directional).

Each finding carries evidence.observed quoting the page verbatim and a runnable verification.reproduce; expected_impact is banded + confidence-tagged (no naked percentages).

Honesty

• Heading keyword stuffing does not buy ranking — write headings for the reader; clarity is the signal, not keyword density. Don't recommend cramming terms into H1/H2.
• Multiple H1s are valid HTML5 and modern parsers/AI tolerate them; the real cost is a muddied outline, so treat as severity 3 directional, not a critical failure. Skipped heading levels are primarily an accessibility/structure concern — frame impact honestly (directional) rather than promising a ranking lift.