GEO Agent - Devpost Submission

Inspiration

As AI assistants like ChatGPT and Claude become the new front door to the internet, a troubling pattern emerged: companies are invisible to AI. When users ask "What's the best project management tool?" or "Which CRM should I use?", AI models confidently recommend competitors — often completely omitting businesses that should be relevant.

We realized this isn't just a visibility problem — it's a trust problem. AI models sometimes hallucinate facts about companies, recommend outdated information, or favor well-known brands over better alternatives. Traditional SEO doesn't help here. There's no "meta tag" for ChatGPT.

The rise of AI-first search (ChatGPT, Perplexity, Claude) means millions of purchasing decisions are being influenced by models that may not even know your company exists. We wanted to shine a light on this blind spot.

What it does

GEO Checker is a multi-agent AI system that audits your website's visibility across leading LLMs:

Scrapes & Analyzes your website to build a comprehensive company profile
Analyzes Your Industry to understand competitive landscape and user intent
Generates Targeted Questions that real users would actually ask AI assistants
Queries Multiple LLMs (GPT-4o and Claude) with web search enabled in parallel
Asks Follow-Up Questions — "Why wasn't [company] mentioned?" — to get direct feedback from the AI itself
Identifies Three Types of Gaps:
- 🔴 Hallucinations — AI says incorrect things about you
- 🟡 Omissions — AI doesn't mention you when it should
- 🟠 Competitor Advantages — AI recommends competitors instead
Generates Copy-Ready Recommendations using proven frameworks (PAS, AIDA, Feature-Benefit)

The result: a detailed report showing exactly how AI perceives your brand and specific text you can add to your website to fix it.

How we built it

Backend (Python + FastAPI):

Async orchestrator coordinating 5 analysis stages
Three specialized Claude-powered agents (Question Generator, Gap Analyzer, Recommendation Generator)
Skills system loading domain expertise from YAML files (SEO, Content Strategy, Copywriting)
Multi-LLM querying with web search (OpenAI gpt-4o-search-preview, Anthropic web_search_20250305 tool)
Server-Sent Events for real-time progress streaming
SQLite for analysis history

Frontend (React 19 + TypeScript):

Framer Motion for smooth animations
Tailwind CSS 4 with editorial design aesthetic
Real-time progress visualization showing each LLM's status
Tabbed dashboard separating Gaps/Feedback from SEO/Copy Recommendations
Dedicated architecture page showcasing the multi-agent system

Key Innovation — Two-Phase LLM Querying:

Phase 1: "What are the best CRM tools for startups?"
Phase 2: "Why wasn't [company-url] mentioned? What would they need to improve?"

This gives us direct, actionable feedback from the AI about what's missing — not just a binary "mentioned/not mentioned."

Challenges we ran into

Web Search Integration: Each LLM provider implements web search differently. OpenAI uses a special model (gpt-4o-search-preview) with web_search_options, while Anthropic uses a tool type (web_search_20250305). We built provider-specific implementations with graceful fallbacks.

Question Quality: Our first attempt generated useless questions like "What does [company] do?" We solved this with multi-phase generation: first analyze the industry, then generate categories, then create specific questions real users would ask.

Parsing Unstructured Feedback: When LLMs explain why they didn't mention a company, they respond in natural language. We engineered structured output formats and robust parsing to extract actionable insights.

Real-Time UX: Analysis takes 60-90 seconds. Rather than show a spinner, we implemented SSE streaming so users see each stage progress, each LLM being queried, and questions appearing in real-time.

Accomplishments that we're proud of

The Two-Phase Query System — Asking LLMs "why" they didn't mention a company yields surprisingly actionable feedback that no other tool provides
Skills-Based Agent Architecture — Modular YAML skills make agents produce expert-level output without massive prompts
Beautiful Real-Time UI — Users watch the analysis unfold: industry analysis, questions generating, LLMs responding, gaps identified
Copy-Ready Recommendations — Not just "improve your content" but actual text snippets ready to paste into your website
The Architecture Page — A detailed, animated showcase of how the multi-agent system works under the hood

What we learned

GEO is the new SEO — Optimizing for AI visibility requires fundamentally different strategies than traditional search engines
LLMs can explain themselves — Simply asking "why didn't you mention X?" yields surprisingly useful, specific feedback
Skills > Giant Prompts — Modular skill files are more maintainable and produce better results than monolithic system prompts
Multi-agent beats single-agent — Specialized agents (question generation, gap analysis, copywriting) outperform one generalist agent
Web search is a game-changer — LLMs with real-time web access give current, verifiable information rather than stale training data

What's next for GEO Agent

More LLM Providers — Add Gemini, Perplexity, Llama, and other models to expand coverage
Automated Monitoring — Weekly/monthly reports tracking how your AI visibility changes over time
One-Click Content Generation — Auto-generate the recommended copy, not just suggestions
Competitive Benchmarking — Compare your AI visibility score against direct competitors
API & CI/CD Integration — Let companies check AI visibility as part of their deployment pipeline
Browser Extension — See AI visibility scores while browsing any website