AIVI — AI Visibility Audit System

What AIVI Measures

Two layers of analysis — structural readiness and actual AI visibility

◈

Structural Layer: 8 Components

Technical readiness for AI search. Each component scored, total normalized to 0–100.

Technical: robots.txt, sitemap, canonical, meta tags
Semantics: headings, internal links, text volume
Entity Graph: schema.org (Organization, FAQ, HowTo)
Answer Readiness: FAQ blocks, definitions, comparisons
Citation Readiness: numbers, prices, guarantees, cases
E-E-A-T: contacts, credentials, reviews, experts

◉

Empirical Layer: Polling 5 AI Models

Branded and topical queries are sent to five AI models in parallel (~75 answers per run), measuring actual visibility.

ChatGPT (OpenAI), Perplexity, Gemini (Google), DeepSeek, Yandex Alice
Brand mentions and domain citations in answers
Share of voice vs competitors and your market rank
Sentiment and entity authority (LLM judge with fallback chain)
Semantic similarity between content and AI answers
Adaptive scale: 70% structure + 30% AI visibility (shifted automatically for crawl-blocked sites)

📊

SEO Metrics & Analytics

Separate diagnostic layer with full picture.

Indexability, title/description lengths, text volume
Schema.org on each crawled page
LLM referrer traffic: ChatGPT, Perplexity, Claude
Deep analytics: Trust Score, Content Depth, Readiness

📋

Reports & Dynamics

Every audit is saved — see how visibility changes after implementation.

Export: HTML, PDF, Markdown, JSON
Spider chart across GEO/AEO/LLMO axes
Position tables per query per provider
Audit history and score dynamics

How the Audit Works

Data Collection

Crawling up to 50 pages: robots, sitemap, markup, texts

Structural Analysis

8 readiness components, scored 0–100

Query Set

Auto-generated from niche, or import your own

AI Polling

ChatGPT, Gemini, DeepSeek and Alice answer queries

Report

Positions, mentions, citations, recommendations

Dynamics

Repeat audits show visibility growth

Market snapshot & 26 mathematical methods — in every audit

Two blocks no Russian competitor has: an LLM-built market snapshot (your niche, actual competitors pulled from real search results, your rank by AI mentions, plain-language action plan) and our in-house mathematical appendix — Kalman filter for score trends, Bayesian visibility intervals, IRT brand strength, kappa agreement across AI models, Gaussian-process forecasts and 21 more, each explained in plain words.

All 26 methods — WhiteGEO in-house development

1. Bootstrap density intervals14. GLMM query×model effects 2. Cohen kappa / Krippendorff alpha15. Conformal prediction intervals 3. Shannon entropy & JSD of visibility16. Matrix completion (soft-impute) 4. Bayesian corpus pooling17. Multi-marginal optimal transport 5. Rasch/IRT brand strength θ18. Non-backtracking link spectra 6. Bradley-Terry mention ratings19. Path signatures of dynamics 7. Kalman score filter20. Tropical (max,+) critical path 8. Gaussian-process forecast21. Knapsack: budgeted action plan 9. CRITIC objective weights22. Ornstein-Uhlenbeck process 10. Random matrices (Marchenko-Pastur)23. Thompson sampling for queries 11. Free probability of spectra24. Differential privacy (ε) 12. RKHS / MMD two-sample test25. Yoneda brand profile 13. Persistent homology (TDA)26. Fisher information geometry

🧭 Market Snapshot

An LLM reads your site to name the niche, picks actual competitors from real SERP data and ranks you by AI mentions: share of voice, leaders, plain-language advice.

⚖️ Adaptive scoring scale

70% structural readiness + 30% actual AI visibility; for crawl-blocked sites the weight shifts to empirical evidence automatically.

🕸️ Site-wide entity graph

schema.org parsed across every page, including @graph containers and microdata; Organization completeness, llms.txt and AI-crawler rules in robots.txt.

✍️ Go content pipeline

Topic summaries → articles driven by a brand passport (USP, audience, prohibitions) → covers via generative models (gpt-image-2, Gemini 2.5 Flash Image) cropped to 1280×720 → Markdown, DOCX export and blog publishing.

📨 Security & transactional mail

Email verification with one-time 256-bit links (only SHA-256 hashes stored), password reset, rate limits and neutral anti-enumeration responses — on our own mail pipeline.

🤖 LLM judge with fallbacks

Mention authority and sentiment scored by a judge model backed by a provider fallback chain — audits never stop on a single API failure.