Giác AI

Human behavior is shifting from "searching on Google" to "finding answer through AI". Giác AI helps Vietnam businesses to reach millions of consumers who are using AI to discover products & brands

Comment

Inspiration

Vietnam's $270B consumer market is rapidly shifting to AI-driven discovery. When a consumer asks ChatGPT "best skincare brand in Vietnam" or Gemini "where to buy electronics in Ho Chi Minh City," the AI's answer becomes the new storefront. But no Vietnamese brand or global brand in Vietnam knows whether they're being recommended, ignored, or misrepresented by these models. SEO has decades of tooling. GEO (Generative Engine Optimization) has nothing, especially for emerging markets like Vietnam where local brands compete against global giants for AI mindshare.

Giác is a Vietnamese word that represents: awareness, perceive, feel, sense - the exact things brands need to be mindful, adapting, updating all the time to keep themselves relevant to the consumer.

What it does

Giác AI is a GEO analytics platform that measures how AI models see Vietnamese Local brands + Global brands targeting Vietnam. It:

  • Queries major LLMs via OpenRouter (For now: GPT-4o, Gemini 2.5, Claude 3.5) with ~1,000 Vietnamese consumer queries across e-commerce, electronics, and beauty
  • Extracts and scores every brand mention using a composite GEO Score weighing organic visibility, recommendation rate, ranking position, sentiment, and Bayesian confidence.
  • Tracks targeted top brands and auto-discovers 2,694 competitors mentioned by AI
  • Computes Brand Trust using a dual-accumulator system
  • Visualizes everything in a real-time dashboard with treemap leaderboards, per-model breakdowns, archetype radar charts, citation drill-downs, and temporal delta tracking

A brand manager can see in seconds: "Shopee ranks #1 on Gemini but #4 on Claude for e-commerce queries mentioning Ho Chi Minh City", and act on it.

How we built it

Backend pipeline (Python):

  • Async query executor hitting OpenRouter API with per-model rate limiting (10/8/5 concurrent calls for GPT/Gemini/Claude)
  • API calls across 5 runs per query per model for statistical reliability via OpenRouter
  • LLM-powered brand extraction using structured Pydantic schemas
  • Fuzzy deduplication with Vietnamese diacritics normalization (e.g., "Thế Giới Di Động" = "The Gioi Di Dong" = "TGDĐ")
  • GEO Score v4 with Bayesian confidence dampening
  • Dated snapshot architecture

Frontend (Next.js 16 + Tailwind v4):

  • D3-style treemap sized by GEO score with green/red delta coloring from real snapshot diffs
  • Brand detail pages with animated score counters, SVG trend charts, 5-axis radar charts, circular progress rings
  • Province-level filtering via citation query text matching (HCM, Hanoi, Da Nang, Can Tho, Hai Phong)
  • Real source URL extraction from LLM responses, showing which websites AI models actually cite
  • Built entirely with Claude Code over a 36 hours sprint

Challenges we ran into

  • Vietnamese text processing: Brand names with diacritics, abbreviations, and romanized spellings required extensive alias mapping. "Fahasa" = "FAHASA" = "pha ha sa" = "Nhà sách Fahasa"
  • Cost optimization: LLMs API are not cheap, first 2 run plus a few test runs cost 150$ spread around 3 platforms with 3 different API keys. OpenRouter came in to save the day, one API key that calls multiple models!
  • 15K API calls on a budget: Careful batching, checkpointing every 5 queries, and resume capability so a network hiccup doesn't waste $3 of API calls
  • Model consistency: The same query returns wildly different brand recommendations across models. Shopee dominates GPT-4o but is less visible on Claude, this inconsistency is the insight, but makes scoring complex
  • Trust archetype scoring: Brand-specific queries ("Is Shopee trustworthy?") needed separate accumulators from generic queries ("best e-commerce Vietnam") to avoid inflating scores
  • Temporal data without a database: Solved with dated snapshot directories and a build script that diffs latest vs. previous - JSON files as the MVP "database"

Accomplishments that we're proud of

  • 2,753 brands scored from 60 targets + 2,694 auto-discovered competitors — the AI surface area is far bigger than any brand realizes
  • 80,667 real citations extracted from LLM responses, each with query context, model, rank, sentiment, and recommendation status
  • GEO Score v4 with Bayesian dampening produces intuitive 0-95 scores that match human judgment of brand strength. To come up with a fair algorithm to calculate AI visibility is was not easy and we're proud this is something that is very hard to replicate
  • Full working product in 36 hours, from zero code to a demo-ready analytics dashboard with real data
  • Model Consistency Index (MCI), a novel metric showing how evenly a brand appears across all AI models, surfacing dangerous single-model dependencies

What we learned

  • AI models have strong, opinionated views about Vietnamese brands, and those views differ significantly across models. This is a real blind spot for marketers
  • The GEO problem is fundamentally different from SEO: there's no "page rank" to reverse-engineer, no crawl to optimize for. The only way to measure visibility is to ask the models directly
  • Vietnamese consumers are already using AI for purchase decisions. The brands that understand their AI presence early will have a massive advantage
  • Claude Code can build production-quality full-stack applications in days, not weeks — the bottleneck was thinking, not coding

What's next for Giác AI

  • Daily automated scanning with cron jobs to build real trend data over weeks/months
  • More models: Perplexity, DeepSeek, Google AI Overview — the "Index 3" becomes the "Index 6+"
  • Actionable recommendations: Not just "your score is low" but "here are the 5 content changes that would improve your Gemini visibility"
  • Giác AI Agent: With all the data we collected, can be repurpose into an insightful and very powerful AI agent that helps business to take actionable items for them directly
  • PostgreSQL migration for proper time-series storage and multi-tenant SaaS
  • Vietnam expansion: Cover food & beverage, travel, fintech, education verticals
  • Regional expansion: Thailand, Indonesia, Philippines — same AI visibility gap exists across Southeast Asia
  • Whitelabel solution-as-a-Service: Agencies can use the Giác techstack to add on to their existing offerings

And most importantly, we're very serious in scaling this project into a startup and are actively looking for mentors, advisors and investors

Built With

Share this project:

Updates