AIV Methodology
The AI Visibility Index (AVI) is a 0–100 composite score that measures how visible a brand is to AI search engines — the category practitioners also call AI share of voice. It answers one question: when a customer asks an AI for a recommendation, does your brand appear? This page is the full public scoring model — the same one every Visora audit runs on. Every metric, weight and threshold is published, so a score can be audited instead of trusted.
The 7 metrics (with their industry-standard names)
- Recall (25%) — the fraction of AI responses that name the brand · industry equivalent: visibility rate
- Breadth (15%) — average recall across each engine (per-engine consistency) · engine coverage
- Depth (15%) — the richness of the mention (capped at 500 characters) · mention depth / citation prominence
- Keyword Reach (15%) — the fraction of distinct queries where the brand appears · query coverage, the metric a fan-out query set measures
- Convergence (10%) — how consistently the engines describe the brand (NLI + Epanechnikov kernel; needs ≥2 mentions) · narrative consistency
- Factuality (10%) — the fraction of canonical facts the AI gets right · accuracy, the inverse of a hallucination rate
- Rank (10%) — position vs competitors (1st = 1.0, decaying by 0.8 per position) · answer position
The formula
AIV = 100 × Σ(weight × metric) ÷ Σ(weight for non-null metrics). Optional metrics (Convergence, Factuality) return None when data is insufficient — their weight redistributes proportionally to the metrics present.
Query methodology and query fan-out
Scans use generic discovery queries — the kind a customer asks without knowing the brand exists: "Best [Category] in [City]", "Recommend a good [Category] in [City]", "Leading [Category] providers in [City]". Brand-centric queries are never used: they guarantee a mention, which measures recall of a question, not visibility. The reason coverage is measured across a query set rather than a single prompt: Google AI Overviews and AI Mode fan one customer question out into a set of related retrieval queries, then answer from the pages those return. A brand that appears for one phrasing but not its siblings is the most common case of invisible — and it cannot be seen by tracking one prompt at a time.
Grounded vs parametric answers
Engines answer in two modes and the fix differs for each. Grounded (retrieval) engines — Perplexity, Google AI Overviews — fetch pages first and cite them, so the levers are crawlability, structured data, and presence on the sources they already cite. Parametric engines answer from model weights with no retrieval, so the levers are entity consistency and the sources those models learned from. Visora records which pathway produced each answer, so a fix list can target the right one.
Engines and score bands
Nine engines are scanned: ChatGPT, Google AI, Gemini, Microsoft Copilot, Perplexity, Doubao, DeepSeek, Qwen, and Kimi. Score bands: 0–30 Invisible · 31–55 Emerging · 56–75 Established · 76–100 Authoritative.
Limitations
- The quick audit is LLM-estimated, not retrieval-grounded — grounded answers are used wherever an engine supports them
- Convergence and Factuality need at least 2 mentions to compute
- English-language query bias
- Scores are time snapshots — AI models update continuously, so every score is a dated observation rather than a permanent state
Citation
Visora. (2026). The AI Visibility Index: A Composite Metric for Brand Visibility in AI Search. Methodology v1.0. Available at https://visoraco.com/research/methodology
Last updated: