AI Visibility Panel Design

Turn accepted cells into a defensible measurement plan. Do not generate prompts or invent precision.

This skill inherits the ethical floor from skills/ETHICS.md. It enforces anti-hallucination, explicit denominators, and decay-aware versioning. Anti-spray and human-send are not applicable.

Inputs

Require:

• measurement charter;
• prompt_architecture.json;
• QA-approved candidates and complete rejection ledger;
• evidence-backed weight inputs, if any;
• run and review budget;
• variance-pilot observations, when available;
• prior panel version and campaign registry, when applicable.

Never inspect target baseline performance during selection.

Select by strata

Use the canonical intent cell as the sampling unit. Variants and repeated runs are nested observations, not extra buyers.

Allocate across:

• proximity band;
• job, journey, and information act;
• ICP/role and locale/language;
• evidence grade/source type;
• measurement lane and surface;
• core (tracked set), rotating (discovery set), sentinel (tripwire), control (false-positive check), and aided (prompted set) partitions.

Select within a stratum by evidence strength, language authenticity, decision relevance, and diversity. Preserve declared minimums or emit a waiver. Do not select by current target strength, weakness, gap size, or campaign desirability.

Separate lanes

• closed_model: no external search/tools/files/RAG/history; fixed system, model/version, and sampling; fresh session.
• retrieval: record required, allowed, or unavailable, whether retrieval ran, queries when exposed, live/cached state, and citation metadata.
• consumer_surface: explicit clean or account archetype, device, locale, history/personalization state; never merge with API rollups.
• campaign_experiment: pre-registered frozen evergreen, unaided resonance, aided association, and matched unaffected controls.

Never mix aided statuses or lanes in a denominator.