---

name: geo-keyword-research2.0 description: Evidence-backed GEO keyword and Prompt Map research for Chinese AI visibility. Use when asked to generate GEO keywords, AI-search prompts, prompt/question coverage, Nanobot prompt-map JSON, or grouped questions for 品类问题、场景问题、品牌问题、竞品比较问题; also use for this GEO diagnostic page's Step 3 prompt analysis.

GEO Keyword Research 2.0

Quick start

Required input: brand or product name. Prefer: website, industry, market, competitors, campaign plans, selected question categories, material/source hints.

Default categories when absent: 品类问题, 竞品比较问题.

Choose one mode:

• frontend-json mode: fast Step 3 response. Return strict JSON for the GEO diagnostic page.
• full-research mode: evidence-heavy research. Use DAG search, source files, fetches, then compile JSON/report.

Use bundled resources:

• Build the search DAG: python3 scripts/build_query_plan.py --input context.json --output query-plan.json
• Search with configured preset: model_preset_call(preset="search", prompt="...")
• Fetch/read source pages with configured preset: model_preset_call(preset="fetch", prompt="...")
• Validate research tool trace: python3 scripts/validate_research_trace.py --trace final_report/brands/{brandId}/research-trace.json --context context.json
• Validate frontend JSON: python3 scripts/validate_prompt_json.py --input result.json --context context.json --strict-count；full-research 严格信源校验再加 --strict-references
• Compile and merge frontend JSON into the brand store: python3 scripts/compile_result.py --prompts prompts.json --sources sources.json --artifact-dir final_report/brands/{brandId} --output final_report/brands/{brandId}/result.json
• URL table template: templates/data-url-template.md
• Category definitions: references/category-taxonomy.md
• Output formats: references/output-schema.md
• Evidence workflow: references/evidence-workflow.md
• Brand/product knowledge base: references/brand-knowledge-base.md

Workflow

1. Normalize inputs.
   • Trim brand/product/competitors/campaigns.
   • Record brand aliases if provided or discovered; final prompts should use the user-facing brand name.
   • Use only current categories: 品类问题, 场景问题, 品牌问题, 竞品比较问题.
   • Never emit old categories: 品牌认知类, 推荐决策类, 竞品对比类, 预算价格类, 落地执行类, 风险合规类.
2. Build a query plan.
   • Use build_query_plan.py for any structured task.
   • In full-research mode, assign one independent dimension per search subtask and write source records.
3. Collect and persist evidence.
   • Must use model_preset_call(preset="search", prompt=...) for query-plan searches; it reads the configured search model preset from nanobot-runtime/config.json (search-combo). Do not use web_search as the primary search path.
   • Must write outputFiles.researchTrace with every search/fetch call: tool, preset, query/url, and discovered source URLs. Validate it with validate_research_trace.py before returning.
   • Final artifacts are brand-level only: use build_query_plan.py outputFiles under final_report/brands/{brandId}/. Treat caller/session task directories as run metadata or scratch only.
   • sources.json, prompt-analysis.json, urls.md, research-trace.json, and brand_knowledge/ are canonical brand-level stores and must be merged in place across repeat runs for the same brand.
   • If model_preset_call(preset="search") returns third-party URLs, those URLs must be normalized into sources.json; do not write only official-site URLs when search found third-party evidence.
   • Full mode: write the human URL table to brand-level outputFiles.urls using templates/data-url-template.md; do not create a separate run-scoped source of truth.
   • Frontend mode targets 20 unique, real, traceable source URLs and should return once useful real sources are available; do not stall just to collect more URLs.
   • Frontend mode is latency-sensitive: once official/customer material, competitor context, a usable source set, and the knowledge base are saved, immediately return JSON instead of continuing补搜.
   • Full mode must collect at least 100 unique, real, traceable source URLs before prompt generation; if sources are weak, continue补搜 rather than inventing URLs.
4. Fetch and synthesize.
   • Use model_preset_call(preset="fetch", prompt=...) for URL/page reading and source summarization; it reads the configured fetch model preset from nanobot-runtime/config.json.
   • Full mode: fetch 8-12 high-value pages with the failure limits in references/evidence-workflow.md.
   • Frontend mode: do not stall on exhaustive fetch; prefer concise source-backed prompts.
5. Build the Markdown brand/product knowledge base.
   • Read customer input, campaign, competitors, material URLs/files, official site, collected sources, and any existing brand-level brand_knowledge/*.md before prompt generation.
   • Generate or update reusable Markdown pages under brand-level outputFiles.knowledgeBaseDir using references/brand-knowledge-base.md.
   • Merge in place: preserve verified existing claims, add new evidence/claims, and mark unresolved conflicts as 待验证 instead of overwriting silently.
   • 08_来源素材索引.md must keep one row per unique URL or uploaded file path.
   • Every important claim must cite a URL, uploaded file name, or mark 待验证; do not invent facts.
6. Generate questions/keywords.
   • Before writing prompts, read the generated brand_knowledge/*.md pages and use that knowledge base to calibrate prompt angles, terminology, scenes, audience, campaign timing, competitor framing, and evidence boundaries.
   • If the knowledge base marks a fact as 待验证, prompts may ask about that uncertainty but must not present it as proven.
   • Frontend JSON: return strict JSON only; no Markdown fences.
   • Full report: use references/output-schema.md Markdown report format.
   • Default count: 10 natural-language prompts per selected category.
   • Each prompt must be one standalone question only; do not combine multiple questions in one prompt or use multiple question marks.
   • Base every prompt angle on customer inputs first: brand/company, product/service, website, positioning, audience, campaign name/date/brief, competitors, AI platforms, existing/analyzed prompts, material URLs/files, and content types.
   • For 品类问题, ask category-choice/recommendation/selection-standard questions and never put the target brand/product name or aliases in prompt.
   • For 场景问题, ask about the product/service's main use cases, audiences, timing, workflows, or campaign scenarios; never put the target brand/product name or aliases in prompt.
   • For 品牌问题, include the target brand/product and focus on specific functions, specific scenarios, audience fit, trust, or public evidence.
   • For 竞品比较问题, include target brand/product plus a named competitor, comparing around the target product's strongest functions, scenarios, audiences, benefits, rights, or differentiators.
7. Validate and compile.
   • Run or mentally apply validate_prompt_json.py before returning frontend JSON.
   • Use compile_result.py when prompt and source artifacts are separate.

Category intent

Keep short definitions here; read references/category-taxonomy.md for examples and anti-patterns.

• 品类问题: non-brand category/category-choice questions. Must omit the target brand/product name in the user-facing prompt.
• 场景问题: brand-neutral use-case/audience/campaign/timing/workflow questions for the product/service's main usage scenarios.
• 品牌问题: direct target-brand questions about specific functions, scenarios, audience fit, trust, evidence, or sourceability.
• 竞品比较问题: target brand vs named competitors, comparing concrete functions, scenarios, audiences, benefits, or differentiators.

Output contract for frontend-json mode

Save and merge JSON artifacts plus Markdown knowledge-base files in the brand-level store, then return exactly this shape:

• final_report/brands/{brandId}/prompt-analysis.json: merged categorized prompts, rationale, evidence/data support, optional supportingSourceUrls[].
• final_report/brands/{brandId}/sources.json: merged searched source records. Frontend mode targets 20 unique URLs but may return fewer real URLs; full-research mode requires at least 100 unique URLs.
• final_report/brands/{brandId}/brand_knowledge/*.md: reusable brand/product knowledge-base pages merged across repeat runs.

{
  "tasks": ["正在搜索官网与第三方来源"],
  "artifacts": {
    "promptAnalysisPath": "final_report/brands/{brandId}/prompt-analysis.json",
    "sourcesPath": "final_report/brands/{brandId}/sources.json",
    "knowledgeBaseDir": "final_report/brands/{brandId}/brand_knowledge",
    "knowledgeBasePaths": ["final_report/brands/{brandId}/brand_knowledge/00_品牌总览.md"]
  },
  "prompts": [
    {
      "id": "p1",
      "prompt": "自然语言问题或 keyword",
      "frameType": "品类问题",
      "priority": "P1",
      "rationale": "为什么值得优先验证",
      "evidenceNeeded": "需要哪些可信证据支持回答"
    }
  ],
  "references": [
    {
      "title": "来源标题",
      "url": "https://source.example.org",
      "source": "来源网站",
      "supportPoint": "该来源支持了什么判断"
    }
  ],
  "coverageSummary": "整体占有率、引用覆盖、竞品覆盖和品牌机会"
}

Quality bar

• Use real, traceable sources; no placeholder URLs.
• Do not promise rankings, traffic, conversion, or final pricing.
• Budget/high-trust topics are diagnostic only; mark medical/financial/legal/education claims for consultant or compliance review.
• If evidence is weak, prefer fewer confident claims over invented confidence.