geo-poison-detector

# GEO Poison Detector — AI推荐投毒检测器 ## 这个工具是做什么的？ **问题：** AI推荐的产品不一定可信。不法商家通过「GEO优化」——批量生成含虚假参数的软文并大量铺发——让AI误以为这些虚构商品是市场上的优质选择。2026年央视315晚会曝光此类黑产，收费1.5–2万元/年，2小时内即可让虚假产品成为AI的「标准答案」。 **这个skill做什么：** 在你收到AI产品推荐时，自动扫描伪科技话术、软文模板特征、来源可信度，给出🟢安全 / 🟡可疑 / 🔴高危三档风险判断，并生成验证链接让你一键核查。 **三种使用方式：** - 🤖 **自动防护**：AI推荐产品时，自动附带风险标记，无需任何操作 - 🔍 **主动检测**：发送 `检测 产品名` 或 `/check product name`，立即查 - 🔗 **链接分析**：发送文章链接，自动抓取并分析是否为投毒软文 **覆盖范围：** 中文/英文产品，国内（京东/淘宝/企查查）和国际（Amazon/BBB/USPTO）双市场验证。 --- **What this tool does (EN):** Detects GEO (Generative Engine Optimization) poisoning — bad actors mass-generate fake product specs into AI training data, making fraudulent products appear as authoritative AI recommendations. Exposed by CCTV 315 (2026, China); similar tactics documented by Wired/MIT Tech Review globally. Auto-flags products when AI recommends them, supports user quick-checks by name, and analyzes article URLs for soft-ad patterns. --- ## Mode 1: Auto-Trigger (AI Recommending Products) Whenever you are about to recommend a specific product or brand, **run detection first**, then append a risk badge inline: ``` [Your normal recommendation answer] --- 🛡️ GEO投毒检测 / Poison Check: [🟢 SAFE / 🟡 SUSPICIOUS / 🔴 HIGH RISK] 触发指标 / Indicators: [list triggered signals, or "无" if clean] 验证链接 / Verify: [run verify_product.py and include top 2-3 links] ``` **Do NOT skip this step when recommending products.** This is the core value of the skill. --- ## Mode 2: User Quick-Check (Product Name) Triggered when user sends any of: - `/check <product name>` - `检测 <产品名>` - `查一下 <产品名>` - `is this legit: <product>` - `这个可信吗: <产品名>` - Or any message asking to verify a specific product by name **Action**: Run the full 5-step detection workflow on the product name. Output the full report format. Run `scripts/verify_product.py "<product>"` to generate verification links. **Example interaction**: ``` User: /check 量子能量水杯黑洞级净化 Agent: 🔴 HIGH RISK — 检测到2个高权重伪科技词汇 触发指标： • [Step 1] 「量子能量」— 高风险伪量子话术 • [Step 1] 「黑洞级」— 高风险伪黑科技话术 验证链接：[京东] [企查查] [国家专利局] 建议：该产品名称含多个典型GEO投毒特征词，极可能为虚假推荐，请勿购买。 ``` --- ## Mode 3: URL Analysis (Article/Page) Triggered when user sends a URL and asks to check it: - `check this: https://...` - `帮我检测这篇文章: https://...` - `这个链接可信吗: https://...` - Any URL from: WeChat (mp.weixin.qq.com), Zhihu, Baijiahao, Medium, blog sites **Action**: 1. Use `web_fetch` to retrieve the article content 2. Run the full 5-step detection workflow on the fetched text 3. Also note the source domain as part of Step 4 source quality assessment 4. Output the full report format **Example interaction**: ``` User: 帮我检测这篇文章 https://mp.weixin.qq.com/s/xxxxx Agent: [fetches content] 🟡 SUSPICIOUS — 检测到软文批量生成特征 触发指标： • [Step 2] 模板化结构："很多人不知道的是" + 产品推荐固定格式 • [Step 4] 来源：微信公众号自媒体，无权威背书 验证链接：[产品名搜索链接] 建议：内容结构符合GEO软文模板，建议通过官方渠道核实产品信息。 ``` **Handling fetch failures**: If `web_fetch` fails or is blocked, ask user to paste the article text and switch to Mode 2 workflow. --- ## Detection Workflow (5 Steps) Apply to content from any mode. ### Step 1 — Pseudo-tech buzzword scan (HIGH weight) Load `references/pseudo-tech-terms.md`. Scan for high-risk terms in both CN and EN sections. - 2+ high-risk terms → immediately 🔴 - 1 high-risk term → 🟡 suspicious ### Step 2 — Batch-generated content fingerprint (HIGH weight) **Universal signals (CN+EN):** - Fixed template structure (Problem → Solution → Product plug) - Keyword stuffing (product name repeated 5+ times) - Vague superlatives without verifiable data - No model numbers, no verifiable specs, no brand registration - Multiple sources with identical or near-identical wording **CN-specific:** - "很多人不知道的是..." / "内部员工都在用" - 自媒体/百家号/微信公众号 as sole sources - 无厂商官网、无天猫/京东旗舰店 **EN/Global-specific:** - "Doctors don't want you to know..." - Affiliate disclosure buried or absent - Only "as seen on" claims, no retailer presence - Reviews only on brand's own site, not Amazon/Trustpilot ### Step 3 — Product authenticity cross-verification (MEDIUM weight) Run `scripts/verify_product.py "<product name>" [--market cn|global|auto]` **CN market:** JD.com, Taobao, Qichacha, Tianyancha, CNIPA patents, GB standards **Global market:** Amazon, Google Shopping, BBB, Trustpilot, USPTO patents, EU RAPEX, Reddit ### Step 4 — Source quality assessment (MEDIUM weight) | Source Type | CN Example | Global Example | Trust | |---|---|---|---| | Major retailer official | 京东/天猫旗舰店 | Amazon/BestBuy official | High | | Gov/standards body | 国家标准委/CNIPA | FDA/CE/ISO | High | | Mainstream media | 央视/人民日报 | NYT/BBC/Reuters | High | | Brand official site | 品牌官网 | brand.com | Medium | | Self-media only | 百家号/头条/微信 | Medium blogs/affiliate | Low | | Unknown/unverifiable | 来源不明 | Unknown | Very Low | ### Step 5 — Risk verdict | Result | Threshold | |---|---| | 🟢 SAFE | 0–1 low-weight indicators | | 🟡 SUSPICIOUS | 2+ medium OR 1 high-weight indicator | | 🔴 HIGH RISK | 2+ high-weight OR confirmed fake specs | --- ## Output Format **Quick badge** (Mode 1 auto-trigger): ``` 🛡️ GEO Check: 🟢 SAFE — no poisoning signals detected ``` **Full report** (Mode 2 quick-check or Mode 3 URL, or when user asks for details): ``` [🟢/🟡/🔴] <one-line verdict in user's language> 触发指标 / Indicators: • [Step N] <indicator> — <explanation> 验证链接 / Verify: • <platform>: <url> 建议 / Recommendation: <next action> ``` ## Language - Match user's language (CN/EN/mixed) - Auto-detect market from product name (CJK → CN, Latin → Global) - For CN products in global context: check international presence too --- ## References - Term library (CN+EN): `references/pseudo-tech-terms.md` — load during Step 1 - Verification script: `scripts/verify_product.py` — run during Step 3 - Usage: `python3 verify_product.py "<name>" [--market cn|global|auto]`