---
document_id: agent.cookbook.product_review_analysis
schema_version: 2
parent_document_id: agent.cookbook
section: cookbook
---

<!-- Generated by scripts.build_agent_guide; do not edit. -->

# E-commerce review queries and analysis / 电商评论查询与分析

Query buyer reviews to understand use cases, strengths, customer pain points, and purchase concerns. Resolve the products and platform-specific filters, pilot when needed, and retrieve individual reviews or the agreed bulk scope. / 查询电商买家评价，了解使用场景、满意之处、痛点和购买顾虑。先对齐商品与平台条件，必要时试抓，再按需要查询或批量采集。

## Business questions / 商业问题

- What do buyers like or dislike about this product?
  这款商品大家怎么评价，满意和不满意什么？
- In which situations do users encounter problems, and what workarounds do they use?
  问题发生在什么使用场景，用户后来怎样解决？
- How do reviews of these products or variants differ?
  这几款商品或不同规格的评价有什么差异？
- What complaints or explicitly stated return reasons appear in negative reviews?
  差评里有哪些具体问题，哪些评论明确提到退货原因？
- Collect reviews for these products over the agreed scope, starting with a small pilot.
  先试抓，再按约定范围采集这些商品的评价。

## Approach / 方法

1. Choose actions as needed, not a mandatory sequence. Reuse product links or IDs. With a name, resolve candidates first; for a brand or category, establish the product set instead of assuming complete coverage. Clarify only identities or conditions that change the query. Compare matching models, variants, or bundles rather than similar names alone.
   按需选用做法，不要求逐项执行。商品链接或 ID 直接复用，只有名称先找候选；品牌或品类口碑先确定商品集合，不默认覆盖全部。只澄清会改变查询的身份或条件，不能仅凭名称相近就把不同型号、规格或套装当同款。
2. Platforms and conditions. Examples include Taobao/Tmall, JD, and Amazon; current authenticated discover determines actual access and inputs. Check the product identifier, country or site, sort, rating, time, and pagination supported by each operation. Do not copy parameters between platforms or invent unavailable filters. Amazon product references must match the selected country; reviews may cover other variants in the same listing.
   适用平台与条件。包括淘宝／天猫、京东、Amazon 等，具体能力与入参以认证 discover 为准。逐操作核对商品标识、国家站点、排序、评分、时间及分页支持，不跨平台照搬参数或虚构筛选。Amazon 商品须对应所选站点，评论可能涉及同一 Listing 的其他变体。
3. Set the collection scope. Follow the requested products, period, ratings, and content. For overall experience, seek positive and negative evidence; for a complaint-only request, do not force an extra positive-review collection. A bulk scope already specified needs no repeated confirmation; otherwise propose a feasible scope and clarify material gaps. Unsupported time or rating filters remain limitations, not assumed capabilities.
   确定采集范围。沿用用户指定的商品、时期、评分和内容。了解整体体验时兼顾正反反馈；只查差评则不强制补好评。批量范围明确就执行，不明确时建议可行范围并确认关键缺口；不支持的时间或评分筛选要说明。
4. Pilot only when needed. Reuse a valid pilot or saved results of the same scope. Check product and variant attribution, review text, dates, template or low-information content, follow-up reviews or seller replies when present, and pagination. Quality screening serves the research goal; do not automatically remove logistics, seller feedback, templates, or short reviews, or silently narrow bulk scope. Missing variant fields remain unknown.
   必要时少量试抓。已有同范围且有效的试抓或结果直接复用。检查商品与规格归属、正文、日期、模板或低信息内容、实际返回的追评或商家回复，以及分页。质量筛选服务于研究目标，不自动删除物流、商家、模板或短评，也不缩小批量范围；缺少规格字段就保留未知。
5. Retrieve and resume to the required extent. Targeted research may deepen selected reviews as needed; bulk collection continues through the agreed reachable pages or until user, budget, or service limits, not merely until themes repeat. Save results, completed identities, and continuation positions in batches. Reuse successes; for an unknown outcome, read the original task before retrying. Record unavailable content and unfinished scope; follow contract approval and recovery rules.
   按需取数与续取。定向研究按问题深入，批量采集按约定取完可达分页或到达用户、预算、服务上限，不因主题重复就停。分批保存结果、已完成身份和续取位置；复用成功项，结果未知先查原任务。记录缺失内容和未完成范围，审批与恢复按契约。
6. Compare on compatible scopes. Align periods, rating filters, sorting, products, and variant mixes across platforms where possible. Keep incompatible samples separate. Deduplicate by platform, product or listing scope, and review ID; retain attribution for shared variant reviews and avoid double-counting them. A follow-up review updates the same experience rather than automatically adding an independent buyer.
   同范围比较与去重。跨平台尽量对齐时期、评分、排序、商品和规格构成，不可比的样本分开保留。按平台、商品或 Listing 范围与评论 ID 去重，共享变体评价保留归属、避免重复计算；追评属于同次体验的补充，不自动算成新增买家。
7. Analyze only when requested. Keep original text and identifiers; distinguish concrete experiences, purchase concerns, low-information judgments, and product versus delivery or service issues. Count with code and distinguish fetched, deduplicated, and analysis-used records. Themes can overlap; state the denominator. Do not infer a broken product from fear of breakage. A data-only request does not require theme coding, a report, or additional metrics.
   需要时再归类统计。保留原话和身份，区分实际经历、购买顾虑、低信息评价，以及商品、物流与服务问题。数量用程序算，分开记录获取量、去重量与分析采用量；主题可重叠，比例须有明确分母。“担心断裂”不能算实际断裂。只要数据时不强制做主题编码、报告或追加指标。

## Boundaries / 边界

- This method concerns e-commerce buyer reviews, not social comments or monthly sales. Platform-displayed review counts, review summaries, and actually retrievable text are different; exhausting an interface does not prove all reviews were obtained.
  本方法针对电商买家评价，不用社媒评论或月度销量替代。平台展示评论数、评价汇总与实际可获取正文是不同口径，接口翻完不等于取得平台全部评价。
- A selected or negative-only sample cannot establish overall defect, return, or dissatisfaction rates. Attribute return reasons only when reviews explicitly mention them; distinguish concerns and hypotheses from observed experiences.
  筛选样本或只看差评，不能估计整体故障率、退货率或不满意率；只有原文明示才归纳退货原因，顾虑和假设不当作实际经历。
- Missing fields, text, dates, or variants remain limitations. Do not infer reviewer demographics, purchase verification, or fake-review status without evidence. Translation may assist reading but retain the original text and provenance.
  字段、正文、日期或规格缺失要保留限制，不无据推断用户画像、购买验证或刷评。翻译可辅助阅读，但保留原文和来源。
