Use cases
- Expand seed words, brands, products, competitors, regions, audiences, and business materials into AI search question sets.
- Establish the question foundation before content production, and output intent clusters, follow-up chains, query rewrites, content topics, FAQs, knowledge base entries, and monitoring Prompts.
- Adapted for DeepSeek, Doubao, Qwen, Kimi, and Yuanbao to domestic AI platforms' colloquial, multi-turn, scenario-based, and complex decision question styles.
Required materials
references/research-foundation.mdreferences/intent-mining-method.mdreferences/cn-platform-adaptation.mdreferences/scoring-and-mapping.mdreferences/real-data-ingestion.mdreferences/report-module-contract.mdreferences/four-format-output.mdreferences/quality-gates.md
Execution process
- Normalize input objects: main object, brand, product line, industry, competitors, region, audience, budget, scenario, pain points, qualifications, scale, timeliness, and material sources.
- Determine the real data mode:
未接入 / 用户提供 / 工具或连接器导入 / 已采样校准. When there is no real data, you must output data gaps and a sampling plan, and must not fabricate real platform answers, search volume, or conversion data. - Perform fact and evidence calibration: brand, product, price, compliance, competitor, and industry facts must be marked with source status; uncalibrated information can only be used as assumptions or items to be confirmed.
- Establish dual-layer intent: first map to the information, navigation/validation, and transaction/action task layers, then expand into nine categories of GEO operational intent.
- Generate complete natural language questions by role, scenario, decision stage, constraints, evidence needs, and content asset use; do not merely pile up short keywords.
- Output five-part rewrites: colloquial question style, standalone rewrite, retrieval phrases, evidence queries, and title input.
- Preserve multi-turn follow-up chains:
root_question_id / parent_question_id / standalone_rewrite / context_dependency / platform_fit。 - Cluster and deduplicate, merging synonymous questions by question objective, user role, constraints, asset use, and compliance level.
- Ten-dimensional scoring: commercial value, AI answer trigger probability, content gap, brand placement space, evidence availability, competition difficulty, dialogue extension value, decision-stage value, platform coverage, and compliance risk. If real data is available, calibrate scores by search volume, AI sampling, customer service/sales frequency, and conversion signals.
- Mapping assets: articles, page modules, FAQs, knowledge base entries, monitoring Prompts, title generation input packs, evidence collection tasks, and 30/60/90-day implementation roadmaps.
- The main report body must cover
references/report-module-contract.mdsystematic modules; project-based reductions are allowed, but the reasons for reduction must be written into the quality check notes. - When outputting Word/PDF/HTML/Markdown, you must comply with
references/four-format-output.md. HTML uses the kami long-document layout language: warm paper background, ink blue, warm gray borders, serif headings, sticky menu; Word wide tables must be landscape A4, fixed table layout, and total table width must not exceed the page's usable width; PDF defaults to landscape A4, and wide tables wrap automatically without overflowing to the right. - Self review: check whether the four reports exist, whether the formats are valid, whether the content is consistent, and whether HTML/Word/PDF comply with the kami layout and tables do not overflow.
Output contract
- AI search question sets and intent maps.
- Question clusters, follow-up chains, and query rewrite lists.
- Scoring matrix and priority ranking.
- Content topics, FAQ banks, and knowledge base entry suggestions.
- Monitoring Prompt library.
- Evidence gaps, compliance boundaries, and implementation roadmaps.
- Real data integration status, AI platform sampling plan, and data calibration actions.
- Default four-piece set: Markdown, HTML, Word, PDF.