How Mirr scores AI brand visibility.

The four-stage pipeline

An audit runs through four stages. Stages 1 and 2 are deterministic: given the same brand name, website URL and competitor list, they return the same numbers every time. Stages 3 and 4 use AI, but operate on the deterministic data from stages 1 and 2 as ground truth, which keeps the final output anchored to facts.

Formulas

Every metric in a Mirr report is one of these:

• Coverage % = (prompts mentioning your brand ÷ 15 total prompts) × 100. Regex word-boundary match against normalised text, case-insensitive. Deterministic.
• Share of Voice % = (your brand mentions ÷ total brand mentions across all responses) × 100. Counts every occurrence of every provided brand name. Deterministic.
• Web Presence Score = Wikipedia (25) + Schema.org (15) + LinkedIn (15) + meta title (10) + meta description (10) + Open Graph (8) + reviews (up to 8) + Twitter card (5) + canonical URL (4). Maximum 100. Deterministic.
• Visibility Score = mean of five category scores (Awareness, Consideration, Decision, Sentiment, Cultural). Each category 0-100, scored by Claude Haiku against a rubric. Ranges: 0-20 Invisible, 20-50 Weak-Emerging, 50-75 Established, 75-100 Dominant.
• Identity Gap = six 0-10 scores for Tone, Values, Audience, Market Position, Brand Promise, Emotional Association. Scored by Claude Sonnet comparing your intended positioning (optional input) to actual AI perception.
• Per-LLM Coverage = same Coverage formula run independently per platform. Reveals whether your visibility is broad (balanced across ChatGPT, Perplexity, Claude, Gemini) or fragile (concentrated in a single corpus).

Why these four LLMs

Between them, ChatGPT, Perplexity, Claude and Gemini account for more than 90 percent of AI-assisted brand discovery today. Each has a different training corpus and update cadence, so a brand that appears in one may be invisible in another. Measuring all four is the only honest way to describe your AI visibility.

• ChatGPT (OpenAI gpt-4o-mini) — largest consumer base, trained on a curated web corpus with periodic refreshes.
• Perplexity (sonar) — research-first, fetches live web, cites sources. Strongest signal for freshness.
• Claude (Haiku 4.5) — used widely by developers and enterprises, high reasoning quality.
• Gemini (1.5 Flash) — integrated with Google Search and Android. Essential for consumer discovery.

Why 15 queries

The 15 queries are generated by a fixed template that asks each brand category five discovery-intent questions (for example, “what are the top tools for X”, “best alternatives to Y”), three comparison questions, three decision-stage questions, and four edge-case or long-tail questions. Running fewer queries produces unstable Coverage numbers; running more produces diminishing returns but tripled cost. Fifteen is the smallest number at which rerun variance drops below five points for most brands.

What varies between runs

Deterministic signals (web presence, coverage counts) are stable to within one point across reruns. AI-scored metrics (Visibility Score, identity gap) have run-to-run variance of roughly five points, which is why we quote the Visibility Score as a band (for example “Emerging”) rather than a single number. The honest number is the band; the specific integer is a directional indicator.

When scores move after changes

Deterministic web presence changes (canonical URLs, Schema.org, meta tags) are reflected the moment crawlers refetch your site, typically within 7 days. Perplexity picks up fresh content in 24 to 48 hours. ChatGPT, Claude and Gemini use training snapshots updated on their own schedules; expect LLM coverage gains to lag deterministic gains by 4 to 8 weeks.

Verifiability

Every audit PDF includes a Raw Responses appendix containing the first eight queries in full, with every LLM response. The numbers in the report are computable from these responses with nothing more than a regex match. If a result looks wrong, you can check the source material directly. This is deliberate: we score brands for AI visibility, so the report itself should be accountable to the same standard of verifiability it measures.