yao-geo-effect-monitor

Yao GEO Effect Monitor

When To Use

• Establish a closed loop for GEO long-term operations covering AI answer monitoring, citation source tracking, brand messaging correction, and performance attribution.
• Output back-end monitoring solutions for customer monthly reports, content iteration, page optimization, external source building, or brand fact correction.
• Cover domestic platforms: DeepSeek, Doubao, Qwen, Kimi, and Yuanbao, and establish independent sampling protocols for each platform.
• Default delivery Word、PDF、HTML、Markdown four-format reports, requiring white background, table borders, alignment, line breaks, and controllable page overflow.

Do Not Use

• One-time GEO strategic diagnosis, prioritize using yao-geo-panorama-audit。
• Only do CRM fields, forms, and conversion tracking, no AI answer monitoring needed, prioritize using yao-geo-tracking。
• Need to bypass platform restrictions, mass abusive scraping, simulate real-user login, or violate terms of service.
• Only need content production or article rewriting, no monitoring closed loop needed.

Required Inputs

Minimum input: brand name, target platforms, monitoring objectives, or reporting period. Optional inputs include intent keyword expansion Prompt library, competitor names, target pages, content inventory, baseline data, publication dates, update records, CRM or conversion data, external source list, historical AI answer samples, device, account, region, network status, and compliance restrictions.

Required Reading

• references/research-basis.md
• references/monitoring-method.md
• references/cn-platform-sampling.md
• references/data-acquisition.md
• references/metrics-attribution.md
• references/correction-loop.md
• references/dashboard-data-model.md
• references/report-completeness-model.md
• references/artifact-layout.md
• references/quality-gates.md

Workflow

1. Unified monitoring objects: brand name, aliases, product names, competitor names, official website domains, WeChat official accounts, documentation sites, media articles, communities, and video accounts.
2. Do an authoritative reference scan: first collect official websites, official documentation, investor/regulatory/standards materials, research papers, and credible third-party sources, establish a source ledger, then proceed to analysis.
3. Do company test scenario discovery: extract product lines, customer scenarios, AI/new features, pricing boundaries, integration ecosystem, Chinese-language material availability, and competitors/substitutes from public facts, then map them to Prompt groups.
4. Establish a Prompt library: organize by 推荐、比较、替代、价格、风险、品牌验证、场景问法 seven groups, each group retaining core questions, long-tail questions, control questions, and follow-up questions.
5. Establish independent sampling protocols for five platforms: DeepSeek records conclusion stability and evidence chains; Doubao records conversational Q&A and image-text output; Qwen records citation sources and follow-ups; Kimi records deep research and long-form citations; Yuanbao records WeChat ecosystem sources and official account content performance.
6. Choose a data access mode: grade by 合成回放、人工真实样本、授权 API/连接器、浏览器辅助合规采样、CRM/转化数据导入 , first confirming permissions, frequency, evidence, and privacy boundaries.
7. Execute sampling and record the environment: platform, time, device, account status, region, network status, Prompt version, original answer text, citation links, screenshots, exported files, API logs, or other auditable evidence.
8. Determine real data availability: without original answers, screenshots/exports, sampling environment, and source records, results must not be labeled as real sampling; they can only be labeled as synthetic samples, inferences, or pending review.
9. Analyze by a six-layer model: covering visibility, factuality, evidentiary support, stability, competitiveness, and governance/attribution; outputting only single-dimension metrics is not allowed.
10. Calculate metrics: brand appearance rate, candidate rate, recommendation rate, ranking, competitor appearance rate, negative statement rate, description accuracy rate, factual error rate, citation recall rate, citation accuracy rate, citation type coverage, answer stability.
11. Do citation source tracking: distinguish official websites, official accounts, media, encyclopedias, communities, video accounts, documentation sites, review sites, aggregator pages, and competitor pages, and determine whether citations support the corresponding claims.
12. Do cautious attribution: set baseline windows, observation windows, treatment Prompts, control Prompts, competitor controls, and external event records; default to using 观察相关, only when the evidence is sufficient should the attribution confidence be increased.
13. Generate a correction closed loop: map erroneous facts, missing evidence, weak pages, negative expressions, and citation gaps to the knowledge base, content transformation, page design, external publishing, or sales messaging.
14. Output a systematic, detailed, and complete report: it must include a source ledger, data access statement, analysis boundaries, metric system, platform differences, citation quality, risk governance, correction roadmap, dashboard/API, and appendix.

Output Contract

• GEO 后端监测方案
• 权威参考与来源账本
• 五平台采样口径
• 真实数据接入模式与证据等级
• 公司测试场景发现表
• 监测 Prompt 库
• 指标体系与计算口径
• 引用源追踪规则
• 答案差异分析与谨慎归因
• 纠偏任务表
• 月报模板
• 阈值告警规则
• 复盘节奏
• 仪表盘字段说明
• 数据库表结构和 API 草案
• 治理、合规、数据质量和置信度说明
• 真实数据可用性判断：可用、部分可用、不可用、合成演示
• 附录：Prompt 全表、指标字典、来源账本、采样记录字段
• 风险、合规与置信度说明
• Default four-piece set:Word、PDF、HTML、Markdown

Validation Checklist

• The four examples or deliverable files actually exist and can be verified by file, text extraction, or opening them in a browser.
• HTML、Word、PDF、Markdown The sources are consistent, and the chapters and key tables are consistent.
• The report has a white background, complete table borders, and column widths, line breaks, line spacing, and page margins do not overflow.
• All five platforms have independent sampling calibers, and conclusions from one platform are not mixed with another.
• Must clearly specify sample_mode and the evidence level; when there is no auditable raw sample, it must not be claimed that real platform data was obtained.
• Do not only look at brand appearance rate; also calculate candidate rate, recommendation rate, description accuracy rate, and citation quality.
• Citation quality checks both recall and accuracy, and does not equate the number of citation links with credibility.
• Attribution must have a baseline, observation window, control prompts, competitor comparison, and records of confounding factors.
• The sampling method does not violate platform terms of service; batch automation must specify frequency, permissions, and the boundaries of manual review.