Why AI search visibility is a new engineering problem
Traditional SEO tooling is built for ten blue links — keyword rankings in Google's results pages. But since 2024, a fast-growing share of searches end in an AI-generated answer: no click, no URL list, no traditional ranking signal. Your brand either appears in that answer or it doesn't. No conventional analytics platform tells you which one.
We built Swenivis GEO because no existing tool gave us an accurate, systematic way to measure whether Swenivis Technologies was being cited correctly across the AI engines our prospects actually use. The data we needed did not exist in any SaaS product we evaluated — so we built the platform ourselves, and the engineering problem turned out to be richer than we expected.
Key pain points
- ✕No tool simultaneously tracks brand citations across ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews
- ✕Each AI engine interprets queries differently — content optimised for one may underperform on another
- ✕Competitor share-of-voice in AI answers is invisible without structured, repeatable measurement
- ✕Website AEO (Answer Engine Optimization) health — schema markup, semantic structure, crawler accessibility — is rarely audited as a unified view
- ✕Traditional SEO dashboards surface keyword rank data that is meaningless for AI answer engine performance
What we built
The platform is a Turborepo monorepo with a NestJS API (port 3001) and a Next.js 14 dashboard (port 3002). The backend uses LangGraph + LangChain to orchestrate multi-step agentic workflows: a brand probe agent submits configurable query sets to each AI engine's API, a citation parser extracts structured mention records (citation type, sentiment, response position), and a competitor analysis agent compares brand vs. competitor share-of-voice across the same query set.
The dashboard is organised around 16 feature modules: AI Visibility scoring, AI Citations log, AI Search Insights, Brand Probe management, Brand Voice analysis, Competitor Analysis, Content Optimizer, Content Performance tracking, AEO Checker, Crawler Audit, Schema Entities, Semantic Structure, Recommendations, Reports, ROI Calculator, and Alerts. Each module is powered by background jobs managed through BullMQ and Redis, with results persisted in PostgreSQL via Prisma and indexed in Typesense for fast in-app search.
Authentication is handled by Clerk, which gave us organisation-level multi-tenancy out of the box — each Swenivis team member or client gets their own workspace with isolated data. PostHog is wired in for product analytics, Sentry for error monitoring, and Svix for outbound webhooks when alerts fire. AWS S3 stores exported reports and crawl snapshots.
Technology stack
Frontend
Backend
AI / Agents
Infrastructure
Key engineering decisions
LangGraph for agentic workflows over direct API calls
Each brand probe involves multiple steps: query construction, API call, response parsing, citation classification, and storage. LangGraph's stateful graph model lets us checkpoint and resume these flows, handle rate limit retries cleanly, and add new AI engines as nodes without rewriting the orchestration logic.
Multi-LLM strategy: Claude for analysis, GPT-4o for structured extraction
We use Anthropic Claude for long-form brand voice and competitive analysis tasks where reasoning quality matters most. GPT-4o with JSON mode handles high-volume structured citation extraction where throughput and output consistency are more important. The two models are treated as interchangeable nodes in the LangGraph pipeline.
Clerk over custom auth for multi-tenant isolation
Building organisation-level data isolation and SSO from scratch would have taken weeks. Clerk gave us per-organisation workspaces, JWT issuance, and Google OAuth in a day. The tradeoff is a third-party dependency for the auth layer — acceptable for an internal tool at this stage.
The results
Swenivis GEO is live and used internally to guide our content and SEO strategy. The platform tells us which topics are generating AI citations, how our brand is positioned relative to competitors in AI-generated answers, and where new content is needed to establish category presence. We use the AEO Checker and Crawler Audit modules to validate our own website before and after publishing changes. We are evaluating making GEO available as a SaaS product.
From our founder
GEO is the most technically ambitious product we have built — it required building a real agentic pipeline, not just a simple API wrapper. We use LangGraph in production here, which means we understand its failure modes, its cost profile, and its operational requirements firsthand. Any client considering an agentic AI build can see exactly how we approached it by talking to us about this product.
Need a custom HRMS or similar product?
We can adapt this codebase to your organisation's requirements, or build from scratch to your specification.
Book a discovery callMore engineering work
