GeoMagicsA number belongs to its context
A visibility number is a property of the engine, market, language and conditions it was taken under — not of the brand. Every measurement carries that context and a reference to the full raw answer.
The AI Growth OS · SEO · GEO · ASO
Know what AI says about you. Then change it.
GeoMagics asks the answer engines the questions your customers ask, measures how you appear with honest intervals, proposes gated fixes for what holds you back, proves what worked — and learns from every outcome.
Early access by request. No number on this page is a measurement.
01 · Where it measures
Every answer engine that decides for your customers.
One adapter interface covers all of them, so the statistics never depend on which engine answered.
- ChatGPTAPI
- ClaudeAPI
- GeminiAPI
- PerplexityAPI
- GrokAPI
- Google AI OverviewsOwn headless capture
- Microsoft CopilotOwn headless capture
- Vertex AI modelsParametric panel
Google AI Overviews and Copilot have no API. GeoMagics captures their live public answer pages with its own headless browser — and fails closed rather than guess.
A separate parametric panel measures what models believe without live retrieval.
No engine answer is ever fabricated: an engine that is not connected returns “not wired”, never a silent estimate.
02 · How it measures
A number is only as honest as its interval.
Engines answer differently every time. GeoMagics treats each answer as one draw from a distribution, and ships every rate with its uncertainty, its sample and its method — or withholds it.
Counting every draw as independent25.7% – 34.7%
400 draws treated as 400 independent observations.
GeoMagics: design-effect-deflated Wilson21.4% – 40.3%
DEFF = 1 + (10 − 1) × 0.40 = 4.60 → 87 effective observations. The honest interval is 2.12× as wide.
Same formula as the shipped estimator (intervals.py, Wilson score on effective n). Every input above is yours; this is a demonstration of the method, not a measurement.
Intervals, from the right estimator
Mention rates use design-effect-deflated Wilson intervals. Sampling design sets prompts and repeats from variance components and power — before a run, not after.
Withheld when it cannot be separated
When the evidence cannot separate two outcomes, the verdict is withheld and the reason is stated. Insufficient is a result, never a zero.
Your change, or the engine's
Every movement gets one of four answers: real improvement, real decline, engine drift or not significant. Windows are compared prompt by prompt, and drift is attributed from an independent anchor panel — never from your own numbers.
Retrieved is not absorbed is not cited
Selection into the engine's sources, mention in the answer and attributed citation are three separate rates, each with its own interval — never conflated into one score.
Corrected for our own classifier
Mention rates are de-biased for the parser's measured error rate against labelled Turkish and English gold sets, and many prompts are tested together under Benjamini–Hochberg false-discovery control.
Signed and verifiable
A measurement can carry a receipt: canonical JSON of the prompt panel, engine, model, draws, window and interval method, signed with Ed25519. Anyone with the public key can verify it was produced by our pipeline and never altered. No key, no receipt.
03 · What it covers
Beyond chat answers.
Search, stores, shopping agents, crawlers and paid AI are measured on the same honest spine, from GeoMagics’ own crawl and index.
AI answers
- Eight answer engines behind one adapter; engines without an API are captured by our own headless browser.
- Prominence by Position-Adjusted Word Count (GEO, KDD 2024) and absorption influence — not just presence.
- Prompt demand estimated from real Wikipedia pageviews.
Search
- Search Console rank history and Core Web Vitals from real-user CrUX field data.
- Internal PageRank authority flow, RFC 9309 robots.txt, JavaScript rendering when a page needs it.
Apps & stores
- App Store and Google Play collectors, metadata-adherence scoring.
- Store ratings as Bayesian credible intervals, not bare stars.
- The app-AEO bridge: from store listing to AI-answer measurement.
Agents & crawlers
- Agent-buyer selection rate — measured on the agent, not a readiness checklist.
- AI-crawler analytics from your own access logs, credited only after IP verification.
Paid AI
- Google AI Max audit: account posture, claim-by-claim copy fidelity against your facts and landing page, migration forensics.
Own crawl & index
- Polite crawling behind an SSRF egress guard, honouring robots.txt.
- Our own off-domain mention and link index, built from the pages we fetch.
Website
Your site and its discovered social channels: AI visibility plus the classic SEO surface.
Mobile app
Store listings and the developer site: AI visibility plus ASO.
Mobile game
Handled as its own sub-type of store listing: AI visibility plus ASO.
Brand
No website required: the brand's whole-web presence, led by off-domain authority and share of voice.
04 · The closed loop
Measuring is the beginning, not the product.
Six steps run end to end on your site, app or brand. The business twin and the intervention-outcome record are what compound.
01
Understand
Each measurement run is bound into a typed business twin, and the brand's own published facts are recorded as claims awaiting approval. Tenants are isolated by row-level security.
02
Perceive
Repeated, contextualised measurement across engines, search, stores, crawlers and agents — with intervals, verdicts and receipts.
03
Decide
The failed pipeline stage is named from evidence; fixes are ranked by marginal gain, with value of information and calibrated forecasts.
04
Act
Content is grounded in your twin's facts and passes factuality, penalty-safety and information-gain gates before publishing through a connector.
05
Prove
Holdout prompts are frozen before the change, then paired difference-in-differences and synthetic control decide whether the movement is yours.
06
Learn
Every intervention and outcome enters an experience record, kept separate per tenant and sector, and exportable by its owner.
05 · Acting safely
Changes you can trust, and undo.
Before and after every change
- A runtime safety envelope is evaluated before any change is staged or applied.
- Autonomy tiers decide what a run may do under your policy.
- Prohibited tactics are blocked and logged; approved actions receive a signed receipt.
- A sequential monitor watches each change and rolls it back automatically on harm.
GeoMe, the analyst
- Speaks only from evidence units drawn from your measurements.
- Answers, asks a clarifying question, or abstains — decided by whether the evidence is sufficient.
- Streams multi-turn analysis and calls real platform tools.
Early access
Tell us what you want measured first.
A website, an app, a game or a brand — and the markets and languages that matter. We reply from publish@geomagics.com.