Every metric, with formula and interpretation.
What each number means, how we compute it, and how to read it.
How often the AI names you.
The fraction of eligible prompts in which a model recommended your brand by name, expressed as a percentage.
RS = recommended / eligible × 100Scale 0–100%. Higher means the AI names you more often among its top suggestions, not merely in passing.
The conservative floor.
The conservative floor of the true Recommendation Share, from the Wilson score interval (z = 1.96 for 95%).
lower = (p̂ + z²/2n − z·√((p̂(1−p̂) + z²/4n)/n)) / (1 + z²/n)Scale 0–100%. Treat this as the worst-case RS you can defend with the sample collected so far.
The optimistic ceiling.
The optimistic ceiling of the true Recommendation Share, from the same Wilson score interval.
upper = (p̂ + z²/2n + z·√((p̂(1−p̂) + z²/4n)/n)) / (1 + z²/n)A narrow band between lower and upper means more certainty; a wide band means more runs are needed.
Real change vs. noise.
McNemar's paired test on week-over-week change, with Edwards continuity correction.
χ² = (|b − c| − 1)² / (b + c)
p = 1 − erf(√(χ²/2))b and c count the prompts that flipped within-prompt between weeks. Alerts fire only when p < 0.05.
How tight the interval is.
How tight the Wilson interval is around RS, mapped to a three-level label.
confidence = max(0, 1 − 2·margin_wilson)
High confidence ≥ 0.7
Medium confidence ≥ 0.4
Low confidence < 0.4High means the number is stable enough to act on. Low means collect more runs before trusting a move.
Run-to-run dispersion.
Run-to-run dispersion of RS within a window, as the coefficient of variation across the k runs.
CV = stddev(RS_k) / mean(RS_k)
Stable CV ≤ 0.2
Moderate CV ≤ 0.5
High CV > 0.5Stable means the model is consistent about you. High volatility means a single run is unreliable — read the interval, not the point.
Your slice of the conversation.
Your share of all brand mentions across you and your tracked competitors.
SoV = brand_mentions / total_mentions_across_competitors × 100Scale 0–100%. A category with few competitors inflates SoV; read it alongside the competitor set, not in isolation.
How the AI frames you.
How positively a model frames your brand when it mentions you. Stored as a raw score from −1.0 to +1.0; on the chart axis it is displayed as −5 to +5 (a ×5 rescale applied for chart width only — the label value stays the raw −1.0..+1.0 score).
stored: −1.0 .. +1.0 (raw)
displayed: −5 .. +5 (chart axis, ×5 rescale)Positive means favorable framing; negative means the model describes you with caveats or criticism.
What sources the model trusts.
How many times a domain appears in the model's cited sources across your tracked prompts. Perplexity only — ChatGPT and Claude do not return citation URLs, so this metric is computed from Perplexity responses only.
count of domain occurrences in Perplexity citations
across all tracked promptsA high-frequency domain that is not yours is an outreach target — it is shaping how Perplexity answers about your category.
How bad the wrong claim is.
An AI-judge classification of how serious a factual violation about your brand is.
HIGH material false claim — fix urgently
MEDIUM partial or misleading claim — reviewOnly HIGH and MEDIUM surface as alerts; lower-grade noise is filtered out.
Can AI crawlers reach you?
Whether AI crawlers (GPTBot, ClaudeBot, PerplexityBot) can reach and have recently fetched your site.
pass crawled within the last 30 days
warning crawled 30–90 days ago
fail crawled >90 days ago, or blocked
unknown no crawl data availableA fail or warning means the models may be answering about you from stale or missing content — fix robots/WAF access first.
What we use, but don't show.
For transparency: each model response is graded 0–10 by a Claude judge for how well it recommends your brand. This score feeds the recommended/eligible decision behind Recommendation Share rather than being displayed on its own.
judge score: 0 .. 10 (internal — drives RS eligibility)You won't see this number in the UI; it shapes the RS you do see.