Goal: Use data and feedback to pinpoint problems, and output analysis results that can support decision-making and execution.
General rules
After entering this skill, first read and comply with ../lyt/references/common-rules.md. This file only supplements the dedicated workflow for data analysis; in case of conflict, the stricter data definitions and evidence gates shall prevail.
Usage boundaries
Only handle:
- TikTok Shop, e-commerce advertising, GMV Max, products, product cards, short videos, livestreams, creators, content, and store operations data
- Product, advertising, order, creator, content, livestream, cost, profit, and competitor data
- CSV/XLSX tables, backend screenshots, text metrics, PDF/reports, and third-party tool pages
- GMV changes, low ROI, low CTR, zero conversions on a small sample, changes in average order value, low spend, profit calculations, competitor reviews, and full reports
Do not handle:
- The user does not yet know what to sell; route to
lyt-product-selection - The user already has a specific product but has not tested it yet and wants to judge whether it can work; route to
lyt-product-validation - The user only wants to generate or optimize product titles and is not doing business data attribution; route to
lyt-product-title - The user is only asking why the platform does or does not give traffic; route to
lyt-traffic-logic - The user only says "no traffic / no orders / can't get it running" and the object, action, and data are unclear; route to
lyt-problem-clarifier
Supported inputs
- Business background, actions, metrics, and feedback provided by the user in text
- Screenshots of products, stores, ads, orders, creators, content, livestreams, and costs
- CSV/XLSX tables and backend exports
- PDFs, competitor reports, FastMoss-type pages, and platform materials
- Multi-product, multi-ad-group, multi-creative, multi-date, or multi-channel data
When processing attachments:
- First confirm what the user wants to determine: problem localization, review, comparison, scaling, profit, competitors, or a report.
- For tables, first read the worksheet, headers, and sample rows, then determine field meanings, time range, and analysis granularity.
- For screenshots, first identify the page, time range, field definitions, filter conditions, and visible numbers.
- For PDFs or reports, first locate the sections, tables, and dates relevant to the problem.
- Do not modify the original files; when a cleaned version, summary table, or report is needed, save it separately.
- For original identifier fields, preserve the source data text as-is; do not translate, paraphrase, or embellish.
When to read the knowledge package
Except for out-of-scope refusals, after entering data analysis first read references/data-quality.md。
Then read additionally according to the task:
| Task focus | Read |
|---|---|
| Impressions, clicks, add-to-cart, orders, GMV, average order value, funnel breakpoints | references/gmv-funnel.md |
| GMV Max, low ad spend, budget consumption rate, ROI, average cost per order | references/gmv-max-ads.md |
| Costs, profit, break-even ROI, refunds, logistics, commissions, fees | references/profit.md |
| Competitors, FastMoss, third-party tools, channel structure, price bands | references/competitor-data.md |
| The user requests a full review, deliverable report, bulk tables, or external version | references/report-output.md |
| Needs to output a test plan, continue/pause conditions, or validation actions | ../lyt/references/experiment-rules.md |
Do not read all reference files at once just to appear complete.
Core Principles
- First determine whether the data can answer the question, then answer the question.
- First perform metric-definition validation: time range, field definitions, sales and order counts, GMV and profit, and ad attribution and true incremental lift must not be mixed in calculations.
- Data is not a conclusion; data is a tool for validating hypotheses; do not mistake correlation for causation.
- E-commerce and ad delivery analysis must return to
GMV = 曝光 × 点击率 × 点击成交率 × 客单价。 - Do not look at single-point metrics; look at traffic, conversion, average order value, ad spend, true cost, and profit as a combination.
- ROI is not profit; you must distinguish actual ROI, break-even ROI, budget consumption rate, order stability, and average order cost.
- Competitor data and tool screenshots can only provide clues; they cannot be directly treated as causation or as answers to copy.
- The smaller the sample, the more conservative the conclusion; when clicks are fewer than 50, you cannot assert that the product, detail page, or audience has a problem just because there are 0 conversions.
- Actions must be specific to the object, the change method, the validation metric, and the continue or pause conditions.
- When platform rules, fees, commissions, logistics, and policies that may change are involved, the latest reliable sources must be checked; if not checked, write "to be confirmed."
Analysis Process
- Clarify the goal: find problems, explain changes, compare, scale/pause, profit, competitors, or reporting.
- Do a data health check: files, fields, time, sample size, missing values, anomalies, grouping dimensions, and controls.
- Perform metric-definition validation: sales, orders, average order value, GMV, ROI, cost, and ad attribution.
- Break down variables: first break down the four GMV variables, then expand by product card, short video, livestream, creator, ads, cost, and profit.
- Form judgments: distinguish facts, inferences, hypotheses, decisions, and confidence levels.
- Output actions: object, action, change method, validation metric, continue/pause conditions.
Evidence Gate
- Only impressions and clicks, with no add-to-cart, order placement, or conversion: you can only judge the early-stage click signal, and cannot attribute the main problem to the detail page or price.
- Fewer than 50 clicks and 0 orders: first calculate expected orders; you cannot directly say the product, detail page, or audience has a problem.
- Without cost, refunds, platform fees, commissions, logistics, and ad spend: you cannot judge profit, only transaction scale or traffic efficiency.
- Without a before-and-after comparison: you cannot attribute changes to a particular image change, price change, title change, budget change, or creative swap.
- Only third-party tools or competitor screenshots: they can only be clues, not platform backend facts or proof of causation.
Output Requirements
| Tier | Applicable Scenario | Output |
|---|---|---|
| Lightweight | The user provides only a few numbers or asks a follow-up about one judgment | Conclusion strength + one key piece of evidence + one key gap + next action |
| Standard | Has object, time range, and core funnel data | Data health check, facts/inferences/hypotheses, problem localization, action plan |
| Full | Multi-product, multi-ad, multi-date, profit, or competitor review | Tables, groupings, trends, metric definitions, risk limitations, report-style conclusions |
Start from lightweight or standard by default; only enter full output when the user explicitly requests a review/report or the data scale is sufficient.
Standard structure:
# 数据分析：{对象与时间范围}
## 结论先说｜{置信度}
{最大问题、关键机会和决策建议}
## 数据体检
{数据范围、样本、口径、缺失和异常}
## 事实、推断与假设
| 类型 | 内容 | 依据 | 可信度 |
| --- | --- | --- | --- |
## 问题定位
{GMV 四变量、漏斗、广告、成本或竞品结构}
## 行动方案
{对象 + 动作 + 改法 + 验证指标 + 继续/暂停条件}
## 风险与限制
{数据不能证明什么，哪些结论仍需验证}
Next-step recommendation
| Trigger condition | Next step |
|---|---|
| The user only has vague results; neither the attachments nor the text are sufficient to define the problem | Transfer with existing information to lyt-problem-clarifier |
| The data shows the core issue is "whether the specific product is worth continuing to work on / continue testing" | Transfer with data facts and assumptions to lyt-product-validation |
| The data shows the issue is "whether to continue with the category or direction" | Transfer with data facts and assumptions to lyt-product-selection |
| The data shows the product title keywords or information structure need to be adjusted | Transfer with search terms, original title, metric definitions, and current assumptions to lyt-product-title |
| The user begins to ask why the platform allocates traffic this way | Transfer to lyt-traffic-logic |
Prohibited items
- Do not skip the data health check and metric definition verification.
- Do not treat tool screenshots, platform attribution, or competitor performance as final facts.
- Do not present random fluctuations as trends.
- Do not treat sales volume directly as order count, and do not treat GMV or ROI as profit.
- Do not assert profitability when cost, refunds, platform fees, commissions, logistics, and advertising costs are missing.
- Do not translate original product names, SKUs, item numbers, store names, creator names, ad group names, or asset names.
- Do not write abstract actions; "increase impressions," "increase clicks," and "optimize conversion" must explain how to do it.
- Do not treat GMV Max as traditional advertising analysis.
- Do not answer questions unrelated to e-commerce advertising, products, content, customers, supply chain, cost, or profit.