I Built an AI Agent That Tells You What Will Sell Like Crazy in Any Country

Ever sat in front of your laptop and thought, "Would this product even work in Brazil?" or "Is there even a market for this gadget in UAE?" Yeah. Me too.

So I built something. An AI agent that does what I call market sense-making. You type in a product category (like "air purifiers" or "wireless earbuds") and a country (like Mexico, Japan, or Brazil), and the agent does all the heavy lifting:

1. Searches on local Google — pulling first- and second-page results to understand what content is ranking.
2. Visits top ecommerce sites in that country — Amazon, Mercado Livre, Flipkart, Noon, whatever’s relevant — and scrapes actual product listings.
3. Estimates product popularity by looking at reviews, ratings, price points, and even special badges like "Amazon’s Choice" or "Best Seller." It’s like market research on-demand without needing to hire a consultant or wait two weeks.

Why Build This?

A lot of small businesses want to go global. But they don’t have the luxury of full-blown category reports from McKinsey. What they need is quick validation:

• Is this product already popular in that country?
• How competitive is it?
• What price range are we looking at?

Most tools are either too generic or cost thousands. I figured I’ll build a scrappy, AI-native version myself.

What the Agent Actually Does

You give it two inputs:

• A product category or sub-category
• A target market or country

And then the agent:

1. Uses the localized Google domain (like google.com.br for Brazil, google.co.jp for Japan) to search for top-ranking content.
2. Scrapes the ecommerce sites like Amazon, Shopee, Mercado Livre, etc., analyzing the first- and second-page listings.
3. Grabs title, price, number of reviews, ratings, badges (like “Best Seller”), and extracts patterns.
4. Estimates sales using a simple rule of thumb: 1 review = ~5–10 purchases. So if something has 100 reviews, you’re likely looking at 500–1000 unit sales. It's rough but useful for directional insights

How I Built It

Here’s what’s under the hood:

• Vibe Coding (Bolt.new) for front-end UI and prompt workflows
• phi.agent + GroqCloud running LLaMA 3 8B for reasoning
• Playwright + BeautifulSoup for browser automation + scraping
• Supabase for storing logs and country/product metadata
• Netlify for frontend deployment

And here’s a code snippet from the core agent logic:

from phi.agent import Agent
from phi.model.groq import Groq
from phi.tools.custom_tools import WebSearchTool, EcommerceScraperTool

agent = Agent(
    model=Groq(id="llama3-8b-8192"),
    tools=[WebSearchTool(), EcommerceScraperTool()],
    description="Suggest best-selling items in a product category and market using local search and ecommerce analysis."
)

agent.print_response(
    "What are the top-selling air purifiers in Brazil? Search Google Brazil and Amazon.br for results, and estimate sales volume.",
    markdown=True,
    stream=True
)

Region-Aware Google Search Tool

from serpapi import GoogleSearch

class WebSearchTool(Tool):
    def run(self, query: str, country_code="br"):
        params = {
            "q": query,
            "google_domain": f"google.com.{country_code}",
            "api_key": os.getenv("SERPAPI_KEY"),
        }
        search = GoogleSearch(params)
        return [r['title'] + " - " + r['link'] for r in search.get_dict().get('organic_results', [])[:10]]

Ecommerce Scraper Tool (Amazon.br Example)

import requests
from bs4 import BeautifulSoup

class EcommerceScraperTool(Tool):
    def run(self, subcategory: str, country: str = "br"):
        url = f"https://www.amazon.com.{country}/s?k={subcategory.replace(' ', '+')}"
        headers = {"User-Agent": "Mozilla/5.0"}
        soup = BeautifulSoup(requests.get(url, headers=headers).text, "html.parser")

        results = []
        for item in soup.select(".s-result-item")[:10]:
            title = item.select_one("h2 span")
            rating = item.select_one(".a-icon-alt")
            reviews = item.select_one(".a-size-base")
            if title and rating:
                results.append(f"{title.text.strip()} | {rating.text.strip()} | Reviews: {reviews.text.strip() if reviews else 'N/A'}")
        return results

Estimating Sales from Reviews

def estimate_sales_from_reviews(reviews):
    try:
        count = int(reviews.replace(",", ""))
        return f"Estimated sales: {count * 7}"
    except:
        return "Sales estimate unavailable"

First Run: Real Examples

🇧🇷 Electronics in Brazil

• Searched "wireless earbuds"
• Amazon.br returned Xiaomi, JBL, and a few local knockoffs
• Items with 300–1000 reviews → ballpark sales: 2k–7k/month

🇯🇵 Kitchen Tools in Japan

• “Silent blender” and “no-splash juicer” kept showing up
• High review counts on compact multi-function units

🇦🇪 Beauty Products in UAE

• Amazon.ae showed lots of sponsored items
• Manual filtering helped find organic high-performers

What’s Next

• Add filters for low-competition, high-demand products
• Integrate with AdWords planner for search volume context
• Let sellers upload a product idea and run a fit check
• Build a dashboard to track trends across countries

Final Thought

This tool doesn’t replace deep market research, but it gets you 80% of the way there in 2 minutes. It’s meant for scrappy founders, sellers, and marketers who want to test fast and launch smart. You can bolt this onto your workflow or turn it into a browser-based SaaS for global product research. Let me know if you want to try it. Happy to open-source the agent or even walk through building your own version.