Numbers you can actually explain.

Headline Found in answers is the prompt-weighted mention rate over successful observations only.

Failed provider calls leave the denominator and show as coverage loss. An unweighted companion (for example 20/48) always sits beside the weighted headline.

Recommendation is stored separately from mention. A brand counts as recommended when it opens a list or heading, or recommendation language applies in its own sentence. Warning language (“avoid”, “wouldn’t recommend”) can override.

Citation metrics use only channels that can return citation evidence. Modes that cannot cite are excluded — not scored as 0%.

Headline SOV uses the client plus configured competitors only. Each entity counts at most once per answer. Directories, trade bodies, publishers and noise are excluded from the competitor set.

A discovered-competitor companion includes only names classified as direct commercial peers.

When brands are listed, order comes from first appearance in the stored answer text among businesses — not scroll depth or fold treatment.

Mean rank when listed = sum of appearance ranks ÷ count of successful observations with a non-null rank. Lower is better; it is not scaled to 0–100.

Scores are calculated per channel first. A customer-facing overall is filled only when every configured provider in the pack is eligible. Incomplete packs show a null overall with a reason — Insight does not renormalise a subset under the same Overall label.

Eligibility today: successful / expected coverage at least 80% and at least min(8, expected) successful observations. Soft incomplete band: 80%–95%. Configured channels use equal weights; that is disclosed policy, not a claim that every surface has equal consumer traffic.

Default prompt weights by intent (overridable per prompt):

• High commercial intent · 1.5
• Discovery · 1.0
• Informational · 0.75
• Branded / competitor prompts · 1.0

Weighted headline rates always keep an unweighted n/N companion.

Provider failures and empty failed answers are excluded from rate denominators and reported as missing coverage. Citation-incapable channels are excluded from citation rates. Coverage is shown as successful / expected alongside each score.

Classic SEO scores stay separate from AI rates. Organic ranking visibility (market-scores-v1.4):

Position points: 1→100, 2→85, 3→75, 4–5→60, 6–10→40, 11–20→20, 21–50→8, 51–100→2, not found→0. Search presence adds discrete weights (organic top 10 40, local pack 25, AI Overview citation 20, featured snippet 10, other feature 5). GSC clicks, impressions, CTR and position, and configured GA4 conversions, remain Google’s metrics.

AI Overview citation on the Google SERP is a search-presence signal. Google AI Mode / LLM answer citations are probe citation metrics. They are labelled separately on purpose — do not conflate them in client reporting.

The customer trend baseline is scheduled pulses (default weekly). Same-day manual re-runs stay visible but are not the primary ±. Insight prefers a prior scheduled pulse at least one cadence gap earlier, otherwise a prior scheduled batch on a different calendar day. Primary deltas are refused when formula version, methodology version, prompt pack, provider pack or competitor-set hash differ, or when coverage falls below the eligibility floor.

On a ~48-answer pack, flipping a few named outcomes can move mention rate by roughly 3–5 points. That is real answer change, not formula noise.

Live stamps for new batches and scores: ai-visibility-v1.5 · ai-scores-v1.5. Each batch also carries an input snapshot hash. Historical batches are not silently rescored when formulas change.

• No single blended AI health or Authority Score
• No sentiment, framing, hallucination or factual-inaccuracy rates
• Not a reproduction of every personalised consumer login session
• No AI post-citation click-through as an Insight composite
• No paid placements inside AI answers
• Platforms currently supported: GPT-4.1, Claude Sonnet 5.5, Gemini 3.6 Flash, Sonar, Google AI Mode, ChatGPT, Gemini. Coverage depends on the site pack and enabled surfaces.

Typical small tracked packs are useful for trends, competitive awareness and agency operations. Calling a pack Decision-Grade evidence for budget or exec strategy still needs larger prompt volume, repeats, and full disclosure of weights and intent mix. Weekly cadence, version stamps, immutable batches and per-provider scores are Decision-Grade practices — not an automatic claim about every dataset size.

Optional standards vocabulary: Insight’s presence and prominence measures align with language in the IAB’s Measuring Visibility in the AI Era (August 2026) without claiming full coverage of every IAB metric.