GEO
Brands get described by AI engines whether they manage it or not. GEO audits and tracks how.
team one internal tool, deployed in production on the agency's internal aws/eks platform. data-vendor names and the meridian model withheld.
AI assistants are becoming the discovery layer. A brand is increasingly known by what ChatGPT, Perplexity, Gemini, and Copilot synthesize about it — not by links — and nobody was watching it happen. The emerging tool category measures visibility (is the brand mentioned), and that metric is already commoditizing.
Build the instrument first. Visibility becomes measurable instead of anecdotal when the underlying data is deterministic, prompt-level observation — which engine, which prompt, which date, mentioned or not, cited or not — rather than someone's screenshot of a chat window.
Data-origin governance as a feature. Every datum carries its origin — imported, derived, LLM-generated, demo — and only imported and derived data is KPI-eligible. LLM-generated analysis is directional and labeled, never a headline number. A measurement tool that mixes observation with model opinion stops being a measurement tool.
One scoring rule, stated plainly. Per prompt: gap = topic share-of-voice × (1 − mention rate). High share-of-voice topics where the brand rarely gets mentioned are the opportunities; the math is simple enough to defend in a client room.
Server-side LLM only. Model access runs through a backend proxy; keys never leave the server. The ingest layer is source-agnostic — vendor exports, a scheduled API fetch on the platform's cron, or demo mode with no external data at all.
Tested like it matters. The app carries roughly 2,400 acceptance criteria backed by a 2,484-test suite — written before features, re-verified after each wave. And when the platform's cron deploys failed opaquely, the fix came from reading the platform's own source code, which turned up a documentation bug in the official deploy guide. Verify against reality, not docs.
Visibility tells you if you appear; it doesn't tell you whether appearing helps. The durable, unclaimed problem is representation correctness — is the brand portrayed the way it defines itself. That productization is specced as multi-tenant software (MERIDIAN): full PRD, agent-readable build guide, normative TypeScript data model. Architected and specced; in pilot. The shipped tool above is the instrument; the model underneath MERIDIAN stays off this page on purpose.
AI answers are the new search results page. If you can't measure how the models describe a brand, you can't manage it.
the insight