The system of record for AI visibility

01The category

GEO — Generative Engine Optimization — is its own discipline.

AI answers now sit above the ten blue links, and each engine grounds its citations on a different index. Optimizing for them is not an SEO footnote. It is a calibrated measurement discipline with published research behind it — and a market that is growing fast.

GEO category today (2025/26): ~$0.0B
Projected by 2034: $0B
CAGR (2025/26–2034): ~0%

Source: Directional estimate (Lokrix AI) — internal projection, not a cited market figure.

02Why now

Each engine cites a different, measurable slice of the web.

GEO is a discipline because the engines are structurally different — and those differences are measurable. A single aggregated "AI score" would be dishonest. Lokrix AI scores each engine separately, on the signals that actually drive it, and reports every result with a confidence interval.

Source: Answer-engine citation-source analyses (2024–2026), summarised in CLAUDE.md §4.

ChatGPT: Bing; ~87% of citations match a Bing top-10 result.
Google AI Overviews: Google; ~54% of citations in top-20; separate from AI Mode.
Google AI Mode: Google; Shares a cited URL with AI Overviews only ~14% of the time — a separate engine.
Perplexity: Google + Bing live; Over-weights Reddit (~47% of top citations) and freshness.
Gemini: Google; Google Search grounding; schema-sensitive.
Grok (beta): Web + X; xAI model; real-time web and X (Twitter) search grounding. Live via a web-search fallback; beta — not yet counted in headline scores.
Claude: ClaudeBot; Training-data grounding; crawl access via ClaudeBot is the primary lever.

03How we built it

Two cooperating subsystems.

There is no official citation-ranking API, and AI answers are non-deterministic. Lokrix AI uses two systems that cooperate: one for ground truth (expensive, accurate), one for scale (cheap, calibrated).

Live · AES

Answer-Engine Simulator

For each prompt, Lokrix AI queries the real engines with web-browsing and grounding enabled, captures the answer and its citation cards, and detects brand mentions (URL match + brand NER + semantic match). This is repeated K times across temperature, persona, and locale — then resolved into a Monte-Carlo presence probability with a 95% confidence interval. Non-determinism is a first-class measurement, not a bug.

K-sample Monte-Carlo per prompt × engine
95% CI on every presence probability
URL + brand NER + semantic citation match

Predictive · Model

Calibrated Scoring Model

A calibrated gradient-boosted tree model (plus a cross-encoder relevance signal) predicts citation probability from page, brand, and web features — without always paying for live engine runs. Calibration (Platt / Isotonic) is evaluated by Brier score, ECE, and AUC. SHAP explanations surface which features drive each score. This model powers the cheap what-if simulation in the optimization engine — unmetered, instant.

Calibrated (Brier score, ECE, AUC evaluated)
SHAP explanations per feature
Powers unmetered what-if simulation

Data honesty as design

We show the confidence interval. That is the whole brand.

Anyone can print a big number. Lokrix AI samples each prompt K times, reports the Monte-Carlo presence probability with its 95% CI, badges whether each figure is live, cached, or predicted by the model, and discloses the locale and persona of every run. Sampled or mock data is clearly marked and never dressed up as a live measurement.

Always show the 95% CI
Badge: live / cached / predicted
Never present a sample as live

04Our principles

The commitments that define the instrument.

These are not aspirations. They are constraints baked into the architecture and into every number we show.

Statistical honesty: Every score is a Monte-Carlo presence probability with a 95% confidence interval. Non-determinism is a first-class measurement — always disclosed, never averaged away or hidden inside a single-sample "score".
Tenant isolation & security: Row-level security enforces workspace boundaries on every database query. SOC 2 readiness, GDPR/CCPA posture, PII minimization, and full audit logs are built in. Enterprise tenants get SSO/SAML/SCIM via WorkOS.
Cost respect: Live engine runs are the dominant cost — multiplicative across prompts, engines, samples, and cadence. Every run is budgeted, rate-limited, cached, and metered against the tenant's credit balance. The predictive model is the default; live runs are deliberate.
Provider ToS respect: Official APIs are used wherever they exist. Where they do not, any scraping is isolated behind a swappable adapter, rate-limited with back-off, and accompanied by explicit legal and ToS posture documentation.

Measure your AI visibility.

Run a free report on your site, then unlock live multi-engine scans, history, and the optimization engine.

Create your free account See how it works

The system of record for AI visibility.