本文从工程视角系统讲解如何用数据管道+AI工具构建真正具有竞争壁垒的亚马逊选品分析体系。包含：评论语义分析、竞争位置数据采集、AI Agent工作流设计，以及完整的Python代码示例。
为什么说大多数AI选品工具没用到核心
当前市场上的AI选品工具停留在两个层次：数据聚合（把公开数据整合成Dashboard）和统计分析（在聚合数据上跑预设算法）。这两个层次解决的是效率问题，不是判断深度问题。
真正的竞争壁垒来自第三个层次：语义理解+自定义分析框架。具体来说，就是把原始的评论文本、搜索结果HTML、广告位分布数据，送入语义理解引擎，提取出比BSR和搜索量更接近市场真相的信号。
这需要三个技术组件：高质量的数据采集API（解决数据获取问题）、语义分析引擎（解决数据理解问题）、信号触发机制（解决时效性问题）。下文逐一展开。
组件一：评论语义分析系统
设计思路
评论是目前亚马逊生态中信噪比最高的用户反馈数据。比搜索量更能反映真实需求状态，比专家意见更有规模代表性。但原始评论文本需要处理才能变成可用的信号。
完整的评论分析pipeline如下：
原始URL列表
→ Reviews Scraper API（批量获取评论JSON）
→ 预处理（清洗、去重、日期过滤）
→ LLM语义分析（痛点分类、价值主题提取）
→ 结构化输出（供选品决策层消费）代码实现
python
import asyncio
import json
from dataclasses import dataclass, field
from typing import Optional
import httpx
import anthropic
PANGOLINFO_API_KEY = "your_api_key_here"
CLAUDE_API_KEY = "your_claude_key_here"
@dataclass
class ReviewBatch:
asin: str
reviews: list[dict]
total_count: int
avg_rating: float
@dataclass
class ReviewAnalysis:
asin: str
pain_points: list[dict] # [{theme: str, frequency: int, severity: float}]
positive_themes: list[dict] # [{theme: str, frequency: int}]
unmet_needs: list[str] # 现有产品未满足的核心需求
user_persona_insights: list[str] # 用户画像洞察
differentiation_opportunities: list[str] # 差异化机会
async def fetch_reviews(asin: str, pages: int = 5) -> ReviewBatch:
"""
通过 Pangolinfo Reviews Scraper API 获取指定 ASIN 的评论
支持分页采集，每页约10条评论
"""
all_reviews = []
async with httpx.AsyncClient(timeout=30) as client:
for page in range(1, pages + 1):
payload = {
"url": f"https://www.amazon.com/product-reviews/{asin}/?pageNumber={page}",
"output_format": "json",
"parse_template": "amazon_reviews",
"geo": {"country": "US"}
}
response = await client.post(
"https://api.pangolinfo.com/v1/scrape",
json=payload,
headers={"Authorization": f"Bearer {PANGOLINFO_API_KEY}"}
)
if response.status_code == 200:
data = response.json()
reviews = data.get("parsed_data", {}).get("reviews", [])
all_reviews.extend(reviews)
await asyncio.sleep(0.5) # 礼貌性延迟
if not all_reviews:
return ReviewBatch(asin=asin, reviews=[], total_count=0, avg_rating=0.0)
avg_rating = sum(r.get("rating", 3) for r in all_reviews) / len(all_reviews)
return ReviewBatch(
asin=asin,
reviews=all_reviews,
total_count=len(all_reviews),
avg_rating=avg_rating
)
async def analyze_reviews_with_ai(batch: ReviewBatch) -> ReviewAnalysis:
"""
用 Claude 对评论批次进行语义分析
聚焦提取：用户痛点、遗漏需求、差异化机会
"""
client = anthropic.Anthropic(api_key=CLAUDE_API_KEY)
# 准备评论文本（按评分分层，优先负面）
negative_reviews = [r for r in batch.reviews if r.get("rating", 5) <= 2]
positive_reviews = [r for r in batch.reviews if r.get("rating", 5) >= 4]
negative_text = "\n---\n".join([
f"Rating: {r.get('rating')}/5\nTitle: {r.get('title', '')}\nBody: {r.get('body', '')[:400]}"
for r in negative_reviews[:30] # 最多30条负面评论
])
positive_text = "\n---\n".join([
f"Rating: {r.get('rating')}/5\nTitle: {r.get('title', '')}\nBody: {r.get('body', '')[:300]}"
for r in positive_reviews[:20]
])
prompt = f"""
Analyze these Amazon product reviews for ASIN {batch.asin}.
NEGATIVE REVIEWS (total: {len(negative_reviews)}):
{negative_text}
POSITIVE REVIEWS (sample of {len(positive_reviews[:20])}):
{positive_text}
Provide a structured analysis in valid JSON with these exact fields:
{{
"pain_points": [
{{"theme": "string", "frequency_pct": float, "severity": "high|medium|low", "specific_examples": ["string"]}}
],
"positive_themes": [
{{"theme": "string", "frequency_pct": float}}
],
"unmet_needs": ["string"],
"user_persona_insights": ["string"],
"differentiation_opportunities": ["string"]
}}
Be specific. Use exact language from reviews where possible. Return ONLY the JSON, no other text.
"""
message = client.messages.create(
model="claude-opus-4-5",
max_tokens=2048,
messages=[{"role": "user", "content": prompt}]
)
try:
analysis_data = json.loads(message.content[0].text)
except json.JSONDecodeError:
# 如果JSON解析失败，返回空结构
analysis_data = {
"pain_points": [],
"positive_themes": [],
"unmet_needs": [],
"user_persona_insights": [],
"differentiation_opportunities": []
}
return ReviewAnalysis(asin=batch.asin, **analysis_data)
async def batch_analyze_category(asin_list: list[str]) -> list[ReviewAnalysis]:
"""
并发分析品类内多个ASIN的评论
控制并发数避免API限速
"""
semaphore = asyncio.Semaphore(5) # 最多5个并发请求
async def analyze_single(asin: str) -> Optional[ReviewAnalysis]:
async with semaphore:
try:
batch = await fetch_reviews(asin, pages=3)
if batch.total_count < 5:
return None
analysis = await analyze_reviews_with_ai(batch)
return analysis
except Exception as e:
print(f"Error analyzing {asin}: {e}")
return None
tasks = [analyze_single(asin) for asin in asin_list]
results = await asyncio.gather(*tasks)
return [r for r in results if r is not None]
def aggregate_category_insights(analyses: list[ReviewAnalysis]) -> dict:
"""
聚合品类级别的分析结论
从多个ASIN的分析中提取共性痛点和差异化机会
"""
from collections import Counter
# 聚合所有痛点
all_pain_points = []
for analysis in analyses:
for pp in analysis.pain_points:
all_pain_points.append(pp["theme"])
pain_point_counter = Counter(all_pain_points)
common_pain_points = pain_point_counter.most_common(5)
# 聚合所有差异化机会
all_opportunities = []
for analysis in analyses:
all_opportunities.extend(analysis.differentiation_opportunities)
# 聚合所有未满足需求
all_unmet_needs = []
for analysis in analyses:
all_unmet_needs.extend(analysis.unmet_needs)
return {
"category_pain_points": common_pain_points,
"differentiation_pool": list(set(all_opportunities)),
"unmet_needs_pool": list(set(all_unmet_needs)),
"asins_analyzed": len(analyses)
}组件二：竞争位置数据采集
SP广告位分布分析
SP广告位的覆盖率是竞争结构分析最核心的数据之一，但这类数据必须实时采集才有意义（广告位是实时竞拍的）。
python
import asyncio
from dataclasses import dataclass
from collections import defaultdict
import httpx
PANGOLINFO_API_KEY = "your_api_key"
@dataclass
class SearchResultSlot:
position: int
asin: str
is_sponsored: bool
slot_type: str # "top_banner" | "inline" | "sidebar" | "organic"
title: str
price: Optional[float]
rating: Optional[float]
review_count: Optional[int]
@dataclass
class KeywordCompetitionProfile:
keyword: str
total_slots: int
sponsored_slots: list[SearchResultSlot]
organic_top10: list[SearchResultSlot]
top_advertisers: list[str] # ASIN列表，按广告位数量排序
sp_concentration: float # 前三大广告主的SP占有率之和
organic_difficulty: float # 0-1，基于头部有机位ASIN的评论数量估算
async def fetch_search_result_page(
keyword: str,
page: int = 1,
zip_code: str = "10001"
) -> dict:
"""
采集亚马逊搜索结果页，包含SP广告位和有机排名
zip_code 参数确保获取特定地区的广告展示（亚马逊广告是地域定向的）
"""
url = f"https://www.amazon.com/s?k={keyword.replace(' ', '+')}&page={page}"
async with httpx.AsyncClient(timeout=30) as client:
response = await client.post(
"https://api.pangolinfo.com/v1/scrape",
json={
"url": url,
"render_js": True, # 需要JS渲染才能加载SP广告
"output_format": "json",
"parse_template": "amazon_search",
"geo": {"country": "US", "zip_code": zip_code}
},
headers={"Authorization": f"Bearer {PANGOLINFO_API_KEY}"}
)
if response.status_code != 200:
return {}
return response.json()
def parse_competition_profile(keyword: str, raw_data: dict) -> KeywordCompetitionProfile:
"""从原始采集数据提取竞争结构分析"""
parsed = raw_data.get("parsed_data", {})
results = parsed.get("results", [])
sponsored_slots = []
organic_slots = []
for i, result in enumerate(results):
slot = SearchResultSlot(
position=i + 1,
asin=result.get("asin", ""),
is_sponsored=result.get("is_sponsored", False),
slot_type=result.get("slot_type", "organic"),
title=result.get("title", ""),
price=result.get("price"),
rating=result.get("rating"),
review_count=result.get("review_count")
)
if slot.is_sponsored:
sponsored_slots.append(slot)
else:
organic_slots.append(slot)
# 计算SP集中度（前三大广告主的占有率）
advertiser_counts = defaultdict(int)
for slot in sponsored_slots:
if slot.asin:
advertiser_counts[slot.asin] += 1
top_advertisers = sorted(advertiser_counts.items(), key=lambda x: x[1], reverse=True)
top_3_count = sum(count for _, count in top_advertisers[:3])
sp_concentration = top_3_count / max(len(sponsored_slots), 1)
# 估算有机排名难度（基于头部有机位的评论数量）
organic_review_counts = [
slot.review_count or 0
for slot in organic_slots[:10]
if slot.review_count
]
avg_top_organic_reviews = (
sum(organic_review_counts) / len(organic_review_counts)
if organic_review_counts else 0
)
organic_difficulty = min(avg_top_organic_reviews / 5000, 1.0) # 归一化到0-1
return KeywordCompetitionProfile(
keyword=keyword,
total_slots=len(results),
sponsored_slots=sponsored_slots,
organic_top10=organic_slots[:10],
top_advertisers=[asin for asin, _ in top_advertisers[:5]],
sp_concentration=sp_concentration,
organic_difficulty=organic_difficulty
)
async def multi_keyword_competition_scan(
keywords: list[str],
days: int = 7
) -> dict[str, list[KeywordCompetitionProfile]]:
"""
多关键词竞争扫描
连续多天采集，观察竞争格局的时间变化
"""
results = defaultdict(list)
for keyword in keywords:
daily_profiles = []
for day in range(days):
raw = await fetch_search_result_page(keyword)
if raw:
profile = parse_competition_profile(keyword, raw)
daily_profiles.append(profile)
await asyncio.sleep(24 * 3600 / days) # 均匀分布在时间轴上（实际使用时应用定时任务替代）
results[keyword] = daily_profiles
return dict(results)组件三：Agent工作流------把分析流程自动化
通过 Pangolinfo Amazon Scraper Skill 接入 MCP 协议，可以让 AI Agent 自主触发数据采集和分析，无需人工干预每个步骤。
python
# Agent工作流的核心设计：信号检测 → 自动触发分析 → 推送决策建议
class MarketSignalDetector:
"""
市场信号检测器
当预设条件满足时，自动触发深度分析
"""
def __init__(self, thresholds: dict):
# thresholds 示例:
# {"search_volume_growth_weekly": 0.15, "new_listings_monthly": 50}
self.thresholds = thresholds
self.triggered_analyses = []
def check_entry_opportunity(self, keyword: str, metrics: dict) -> bool:
"""检测是否出现入场机会信号"""
conditions = [
# 搜索量周增长超过阈值
metrics.get("search_volume_growth_weekly", 0) >= self.thresholds["search_volume_growth_weekly"],
# 新品上架速度低于威胁阈值
metrics.get("new_listings_monthly", 999) <= self.thresholds["new_listings_monthly"],
# 头部广告主SP集中度低（说明市场尚未被垄断）
metrics.get("sp_concentration", 1.0) <= 0.6,
]
# 至少满足两个条件才触发
return sum(conditions) >= 2
def check_competitor_weakness(self, competitor_asin: str, recent_reviews: list) -> bool:
"""检测竞品是否出现负面爆发（潜在市场机会）"""
if len(recent_reviews) < 10:
return False
# 最近30天的评论平均评分
recent_avg = sum(r.get("rating", 3) for r in recent_reviews[-30:]) / 30
# 历史评分基线
historical_avg = sum(r.get("rating", 3) for r in recent_reviews[:-30]) / max(len(recent_reviews) - 30, 1)
# 评分下降超过0.5分视为负面爆发
return (historical_avg - recent_avg) > 0.5
# 使用示例
async def run_monitoring_cycle(watchlist: list[str]) -> list[dict]:
"""
定时运行的监控周期
对 watchlist 中的关键词执行完整的信号检测
"""
detector = MarketSignalDetector(thresholds={
"search_volume_growth_weekly": 0.12,
"new_listings_monthly": 40,
})
alerts = []
for keyword in watchlist:
# 获取最新搜索结果
raw_data = await fetch_search_result_page(keyword)
profile = parse_competition_profile(keyword, raw_data)
# 这里简化了 metrics 的来源，实际需要有历史数据作为基线
mock_metrics = {
"search_volume_growth_weekly": 0.18, # 模拟数据
"new_listings_monthly": 32,
"sp_concentration": profile.sp_concentration
}
if detector.check_entry_opportunity(keyword, mock_metrics):
alerts.append({
"type": "entry_opportunity",
"keyword": keyword,
"sp_concentration": profile.sp_concentration,
"organic_difficulty": profile.organic_difficulty,
"top_advertisers": profile.top_advertisers,
"recommended_action": "initiate_deep_analysis"
})
return alerts三层分析结论输出示例
python
def generate_sourcing_decision_brief(
category: str,
review_insights: dict,
competition_profiles: list[KeywordCompetitionProfile]
) -> str:
"""
生成结构化选品决策摘要
供选品团队直接用于决策
"""
# 提取关键指标
avg_sp_concentration = sum(p.sp_concentration for p in competition_profiles) / len(competition_profiles)
avg_organic_difficulty = sum(p.organic_difficulty for p in competition_profiles) / len(competition_profiles)
top_pain_points = review_insights.get("category_pain_points", [])[:3]
opportunities = review_insights.get("differentiation_pool", [])[:3]
brief = f"""
====== 选品决策摘要：{category} ======
📊 竞争结构评估
- SP广告集中度：{avg_sp_concentration:.1%}（{'高' if avg_sp_concentration > 0.6 else '中' if avg_sp_concentration > 0.4 else '低'}）
- 有机排名难度：{avg_organic_difficulty:.1%}（{'高' if avg_organic_difficulty > 0.7 else '中' if avg_organic_difficulty > 0.4 else '低'}）
😤 用户核心痛点（频率排序）
{chr(10).join(f' {i+1}. {pp[0]} ({pp[1]}个ASIN提及)' for i, pp in enumerate(top_pain_points))}
💡 差异化切入机会
{chr(10).join(f' • {opp}' for opp in opportunities)}
🎯 入场决策建议
{'建议入场' if avg_sp_concentration < 0.65 and avg_organic_difficulty < 0.7 else '谨慎评估'}
- 差异化方向：{opportunities[0] if opportunities else '待进一步分析'}
- 评论获取目标：前60天内完成首批高质量评论积累（建议目标50+条）
- 广告策略参考：{'优先长尾精准词' if avg_sp_concentration > 0.5 else '可考虑适度的宽泛匹配'}
"""
return brief总结
构建真正有竞争壁垒的AI选品系统，需要三个层次的工程工作：
| 层次 | 核心工具 | 输出 |
|---|---|---|
| 数据采集 | Pangolinfo Scrape API + Reviews Scraper API | 评论JSON、搜索结果页、广告位分布 |
| 语义分析 | Claude/GPT 等LLM | 痛点分类、差异化机会、用户洞察 |
| 信号自动化 | Agent工作流 + Amazon Scraper Skill | 实时机会警报、触发式深度分析 |
这套系统的竞争壁垒来源：你的分析频率（实时）远高于依赖平台订阅工具的竞争对手，你的分析维度（自定义语义分析）比通用算法更深，你的判断框架内嵌了自己对市场的独特认知。
Tags: #Python #亚马逊选品 #AI选品 #数据分析 #爬虫 #LLM #Agent #API #电商