ai-visibility-basic-checks

Auditing pages with the 36 basic AI Visibility checks

This skill governs how to apply the 36 BASIC AI Visibility checks shipped with ai-visibility-booster. The runtime scope is exactly 36 checks, identified by public IDs AIVB-001 through AIVB-036. Source IDs (Sxxx) are provenance metadata only.

When to use

• Auditing a single page or a sampled batch of pages from a crawl run.
• Reviewing NEEDS_REVIEW verdicts produced by deterministic/heuristic evaluators in verdicts.json.
• Helping a user understand which fixes a check implies.

Verdict vocabulary

Every per-page-per-check verdict is one of:

• PASS — evidence shows the check is satisfied. Always cite the evidence string from the artifact; never invent it.
• FAIL — evidence shows the check is violated. Always cite evidence.
• N/A — the check does not apply to this page (with naReason of not_applicable, pending_access, or missing_data). Explain why.
• NEEDS_REVIEW — the deterministic/heuristic evaluator cannot conclude. Use available evidence to push toward PASS/FAIL/N/A only when the evidence is sufficient. Otherwise leave it as NEEDS_REVIEW and explicitly state what is missing.

Evaluation modes

• deterministic — outputs are reproducible from the page snapshot alone. Do not override unless you have a stronger signal from the snapshot itself.
• heuristic — outputs are derived from rules with thresholds. You may override with explicit evidence; explain the override.
• needs-claude-review — these checks (e.g. AIVB-012, AIVB-015, AIVB-020, AIVB-023, AIVB-028, AIVB-029, AIVB-033, AIVB-036) start as NEEDS_REVIEW. Promote to PASS/FAIL/N/A only with cited evidence from the snapshot.

Review protocol

1. Open the latest verdicts.json for the analyzed domain.
2. Filter to NEEDS_REVIEW items, then sort by severity (Critical → Low).
3. For each item, read the page snapshot under crawl-runs/<id>/pages/<hash>.json and look for the specific signals the check requires (e.g. for AIVB-012: visible byline, Person JSON-LD, author field on Article JSON-LD).
4. If you can confirm PASS, write a short, concrete evidence string.
5. If you must flip to FAIL, cite the missing signal explicitly.
6. If still unsure, leave NEEDS_REVIEW and document what evidence would resolve it.
7. Save updated verdicts under a claudeReview key — do not overwrite the deterministic/heuristic verdict.

Minimal-fix family

Recommendations should always be minimal and extractive:

• add a TL;DR or key-takeaways block without changing meaning;
• split overly long sentences;
• convert existing prose lists into bullets/tables;
• add visible reviewed/published dates;
• add a visible byline and Person JSON-LD;
• add image captions / figcaption companions;
• add or correct schema (Article, BreadcrumbList, VideoObject);
• add an og:image or schema image if missing;
• align title / H1 / og:title / schema headline;
• add prompt-mapped FAQ blocks for strategic pages.

Red flags — never do these

• Recommend rewriting the whole page or replacing meaning.
• Mark PASS without quoting evidence.
• Treat missing signals as automatic FAIL without explaining why.
• Reduce the runtime scope to "25 checks" (the runtime scope is 36).
• Use Sxxx as a public/user-facing ID. The public ID is AIVB-NNN.
• Promise AI citations, AI Overview placements, or Semrush AI Visibility growth.

Output style

When writing review output, prefix each verdict line with the public ID:

AIVB-012: PASS — author "John Reviewer" appears in JSON-LD Article.author.name and as a visible byline ("Reviewed 2026-04-20 by John Reviewer").

If a source ID is needed for traceability, append it parenthetically: (provenance: S037). Never use it as the lead identifier.