amazon-category-research

Amazon Category In-Depth Research

An end-to-end autonomous research workflow for any Amazon Best Sellers category. The user provides a category URL; the skill collects data, runs parallel analyses, and generates a comprehensive Chinese report package.

Prerequisites

• Playwright MCP server must be available (browser_navigate, browser_evaluate, browser_snapshot, browser_close)
• Agent tool used to launch parallel analysis subagents
• Reports are saved to {project_root}/reports/{category-slug}/(for example reports/toys-and-games/、reports/building-toys/). Create the subdirectory at the start of each run.

Workflow Overview

The skill runs in three phases: data collection, parallel analysis, and report synthesis. Use TaskCreate/TaskUpdate to track progress.

Important: All file outputs (JSON data, analysis reports, synthesis reports) must go into the reports/ category-specific subfolder. Extract a human-readable slug from the category name (for example "Building & Construction Toys" → building-toys，"Toys & Games" → toys-and-games). Create the directory before saving any files.

---

Phase 1: Data Collection

Step 1.1 — Parse user input

Extract from the user message:

• Category URL(required) — Amazon Best Sellers URL, for example https://www.amazon.com/gp/bestsellers/toys-and-games/
• Subcategories(optional) — specific subcategories for in-depth research. If omitted, automatically select the Top 5 most relevant to the main Top 30 list.
• Report language(optional, default: Simplified Chinese)

Step 1.2 — Scrape the main category Top 30

Use mcp__playwright__browser_navigate to navigate to the category URL. Wait for the page to load completely.

Use mcp__playwright__browser_evaluate to run this extraction script:

() => {
  const lines = document.body.innerText.split('\n');
  const products = [];
  let current = null;
  for (let i = 0; i < lines.length; i++) {
    const match = lines[i].match(/^#(\d+)$/);
    if (match) {
      if (current && current.title) products.push(current);
      current = { rank: parseInt(match[1]) };
    } else if (current) {
      if (lines[i].includes('out of 5 stars')) current.stars = lines[i].trim();
      else if (lines[i].includes('offers from') || lines[i].match(/^\$[\d,.]+$/)) current.price = lines[i].trim();
      else if (lines[i].match(/^[\d,]+$/) && current.stars) current.reviews = lines[i].trim();
      else if (lines[i].trim().length > 15 &&
               !lines[i].includes('Movers') && !lines[i].includes('New Releases') &&
               !current.title) {
        current.title = lines[i].trim();
      }
    }
  }
  if (current && current.title) products.push(current);
  const nrIdx = products.findIndex(p => p.title && (p.title.includes('New Releases') || p.title.includes('Movers')));
  const result = nrIdx > 0 ? products.slice(0, nrIdx) : products.filter(p => p.title && p.title.length > 10);
  return result.slice(0, 30);
}

Save the results to reports/{category-slug}/{category-slug}_top30.json(using a human-readable slug such as building-toys）。

Step 1.3 — Discover subcategories

On the same page, use mcp__playwright__browser_evaluate to find subcategory navigation links:

() => {
  const allLists = document.querySelectorAll('ul, ol');
  const subcats = [];
  allLists.forEach(list => {
    const items = list.querySelectorAll('li');
    const firstText = items[0]?.textContent?.trim() || '';
    if (firstText.includes('Action & Toy Figures') || firstText.includes('Arts & Crafts') || firstText.includes('Baby & Toddler') || firstText.includes('Building')) {
      list.querySelectorAll('a').forEach(a => {
        subcats.push({ name: a.textContent.trim(), url: a.href });
      });
    }
  });
  return subcats;
}

This returns the full list of subcategory links from the left sidebar. If it fails (returns empty), try alternative selectors:a[href*="/zgbs/"], or extract from the "Any Department" dropdown menu text.

If the user specified subcategories, match by name. Otherwise, use the following logic to automatically select 5 subcategories：

1. Map each Top 30 product to its most likely subcategory
2. Sort subcategories by frequency of appearance in the Top 30
3. Select the top 5 (excluding the main category itself)
4. If fewer than 5 appear, fill in with adjacent categories

Step 1.4 — Scrape the Top 20 of each subcategory

For each selected subcategory:

1. Use mcp__playwright__browser_navigate to navigate to its URL
2. Run the same extraction script as in Step 1.2 (but with .slice(0, 20) instead of 30)
3. Save to reports/{category-slug}/{subcategory_slug}_top20.json

Execute sequentially (Playwright uses a single browser tab). Locate at most 5 subcategories.

Step 1.5 — Close the browser

After all scraping is complete, run mcp__playwright__browser_close。

---

Phase 2: Parallel Analysis

Use run_in_background: true and subagent_type: "general-purpose" to simultaneously launch 4 analysis agents. Each agent receives the complete raw data and writes its report to reports/。

Agent 1 — Niche Market Analysis

File:reports/{category-slug}/{category-slug}_niche_analysis.md

Prompt the agent to analyze:

1. Product Type Distribution — Which categories dominate the Top 30, percentage breakdown
2. Price Band Analysis — $0-7 / $7-15 / $15-30 / $30-50 / $50+ distribution and feasibility
3. Star Rating Patterns — Identify quality gaps (high demand + low ratings = opportunity)
4. Top 5 Niche Recommendations — Ranked by: demand level, competition intensity, profit potential, entry difficulty
5. Supply Chain Guidance — Sourcing locations (Yiwu, Shantou, Dongguan, etc.), factory types, MOQ expectations
6. Risk Assessment — Compliance, seasonality, intellectual property, competition risk

Agent 2 — Competitor Analysis

File:reports/{category-slug}/{category-slug}_competitor_analysis.md

Prompt the agent to analyze:

1. Brand Landscape — Identify repeat brands, brand concentration, China sellers vs US sellers ratio
2. Product Strategy — SKU matrix/variation tactics, bundling, IP licensing, pricing strategy
3. Review Barriers — Estimate review thresholds for new seller competition by price tier
4. Entry Difficulty Rating(1-5 stars) for each major subcategory
5. Top 3 Entry Recommendations including target price, differentiation angle, and risk level
6. Review Building Strategy — Vine program, request a review, social seeding timeline

Agent 3 — Keyword Analysis

File:reports/{category-slug}/{category-slug}_keyword_analysis.md

Prompt the agent to analyze:

1. High-Frequency Keywords — Extract from all product titles, sort by frequency, group by subcategory
2. Search Intent Analysis — By age group (0-2, 2-4, 4-8, 8-12), by scenario (birthday, education, travel, holiday)
3. Keyword Strategy — High-volume low-competition opportunities, title formula for each subcategory, backend keyword template (250 bytes)
4. PPC Keyword Matrix — For the top 4 subcategories: broad/phrase/exact match recommendations + negative keywords
5. Seasonal Keyword Calendar — Monthly plan for Q2 (summer), Q3 (back to school), Q4 (holidays)

Agent 4 — Profit Analysis

File:reports/{category-slug}/{category-slug}_profit_analysis.md

Prompt the agent to analyze:

1. Price Tier Profit Model — For each price band ($0-7, $7-15, $15-30, $30-50, $50+): estimate COGS from Alibaba, FBA fees (15% referral fee, fulfillment fee, storage fee), net margin range
2. Unit Economics per Category — For the top 5 subcategories: estimated COGS, FBA fees, net profit per unit, margin %
3. FBA Optimization — Packaging size tier reduction strategies, storage tips
4. Launch Cost Model — Minimum viable launch budget (2 SKUs x 500 units), advertising budget recommendations
5. Break-Even Analysis — monthly unit sales required to cover fixed costs
6. Return on investment ranking — rank subcategories by estimated return on investment

Important: pass the actually scraped data (not a summary) to each agent. Include all product titles, prices, ratings, and review counts. Agents need real data to produce meaningful analysis.

---

Phase 3: Report Synthesis

Wait for all 4 agents to complete (you will receive task notifications). Then launch a final synthesis agent (subagent_type: "general-purpose", this time running in the foreground because we need the results) to generate the synthesis report.

Final Report

File:reports/{category-slug}/{category-slug}_comprehensive_report.md

The synthesis agent should:

1. From reports/ read all 4 analysis reports
2. Combine the raw product data
3. Generate a synthesis report with the following exact structure:

# 亚马逊美国站 {Category Name} 品类综合研究报告

## 目录
1. 执行摘要
2. 品类全景概览
3. 子品类深度分析 (Deep dive per subcategory)
4. 市场机会矩阵
5. 新卖家入场策略
6. 运营落地指南
7. 财务模型
8. 风险管控
9. 行动路线图

Key sections that must be included:

• Executive Summary: 1-page overview, including a Top 5 opportunities table (niche, rating, profit, startup cost, payback period)
• Category Landscape: Top 30 product type distribution table, price band analysis by FBA fee ratio
• Subcategory Deep Dive: for each scraped subcategory: market characteristics, competitive landscape, pricing analysis, key players, entry difficulty rating (1-5 stars)
• Market Opportunity Matrix: rate 8-12 niches across 4 dimensions (demand, competition, profit, difficulty), assign S/A/B/C grades
• Entry Strategy: 3 phases — validation phase (1-2 months), expansion phase (3-6 months), brand building phase (6-12 months)
• Operations Playbook: Listing title formula, keyword strategy, phased PPC plan, review building, seasonal calendar
• Financial Model: two startup budgets (lean version $3-5K, standard version $8-15K), unit economics for the first 3 products, break-even analysis, 6-month P&L forecast
• Risk Management: compliance checklist (applicable CPC/CPSIA/ASTM/FCC), intellectual property protection, seasonal inventory, competitive defense
• Action Roadmap: detailed 30/60/90-day plan with milestones

---

Output Summary

After all phases are complete, present to the user:

1. A summary table of all generated reports and their file paths
2. The Top 5 key findings from the research
3. The Top 3 recommended action items

Error Handling

• If Playwright navigation fails, retry once, then try mcp__web_reader__webReader as an alternative
• If the extraction script returns empty results, the page structure may be different — use browser_snapshot to inspect the page structure and adjust the selectors
• If subcategory discovery fails, ask the user to provide subcategory URLs
• If any analysis agent fails, regenerate that report