Keyword Configuration Analysis
Use this skill to turn keyword lists, ad reports, backend configs, or spreadsheet data into executable keyword updates, Search Term drafts, campaign layers, negative-word lists, or performance conclusions.
Workflow
1. Confirm platform, scope, and deliverable before finalizing
- Confirm at least: platform/site, product or business object, keyword purpose, analysis window, input source, output format, and whether the task is read-only analysis or config editing.
- Concrete conventions from evidence:
- For 小红书聚光 analysis, respect the user's module boundary: if they say only analyze
聚光数据汇总for7.9-8.3weekly data, do not mix in聚光定向消耗for the same date range until they ask for targeting analysis. - For 聚光 reporting, if the user says
cpuvis the main KPI, rank optimizations by lowercpuvfirst; use other fields as diagnostics, not the primary score. - For Amazon Search Term work, confirm site/language and whether the field is backend
Search Termrather than title, bullets, description, or product highlights. Search Term is for supplemental traffic terms that cannot fit front-end copy, not for repeating every visible term. - For Amazon Italy ads, confirm target language and product sizing before accepting keywords: a Montessori bed with
120×200and140×200should not automatically negatematrimoniale, because140×200can map to small double-bed searches.
- For 小红书聚光 analysis, respect the user's module boundary: if they say only analyze
2. Read structure and locate the authoritative fields
- Inspect workbook sheets, screenshots, table headers, config variables, and current keyword arrays before extracting or editing.
- Concrete conventions from evidence:
- In 聚光 keyword reports, expected analysis columns include
搜索主题,消费,展现量,点击量,ctr,cpc,cpm,进店uv,转化率,cpuv, and小红星任务期阅读UV; preserve rows where computed columns display#DIV/0!as zero-denominator cases rather than treating them as text metrics. - In 聚光 keyword-planning data, the comparable planning fields are
关键词,推荐理由,竞争指数,月搜索指数, and市场出价; cross-check these against actual delivery fields by exact keyword text. - In keyword collection scripts, update only the target keyword list such as the war-news list. If there are separate AI and war keyword lists, keep AI keywords unchanged when the user only says
美伊战争类关键词仅保留.... - In Amazon listing/Search Term work, read
SKU主数据表.xlsx, product selling-point/spec sheets, and keyword files such as花瓶关键词.xlsxtogether; compare candidate backend terms against front-end fields:标题,商品亮点,五点说明, and产品描述.
- In 聚光 keyword reports, expected analysis columns include
3. Clean keywords and metrics with explicit formulas
- Normalize keyword text, language, currency, percentages, and invalid spreadsheet values before calculating or comparing.
- Concrete conventions from evidence:
- 聚光
CTRis点击量 / 展现量; when展现量=0, report it as zero-denominator or#DIV/0!, not as a real 0% performance signal. - 聚光
CPCis消费 / 点击量; when点击量=0, do not rank the row by CPC because the source may show#DIV/0!. - 聚光
转化率is进店uv / 点击量; this is the rate used in evidence where华伦料理棒had high conversion because many clicks became进店uv. - 聚光
cpuvis消费 / 进店uv; lower is better. When进店uv=0, markcpuvas undefined/zero-conversion instead of using infinity in a ranked table. - When checking planning-vs-actual cost, compute actual-to-planning CPC ratio from actual
cpcdivided by planning市场出价; do not compare planning市场出价tocpuv.
- 聚光
4. Filter by relevance, language, and campaign intent
- Separate core, expansion, test, negative, and excluded terms by product fit and intent before looking only at volume or bid price.
- Concrete conventions from evidence:
- For 华伦料理棒,
料理棒,华伦料理棒, and料理棒推荐are direct category/brand terms;辅食工具-type terms are broader and need actualcpuv/进店uvvalidation before scaling. - For 华伦料理棒, competitor terms like
博朗料理棒are not core brand terms; if used, label them as competitor conquest/test words and explain the risk and creative angle. - For Amazon vase Search Terms, avoid repeating front-end-covered terms when the title/bullets/description already contain them, such as
Ceramic,Pottery Vase,Watertight,Modern Coastal,housewarming,wedding,holiday,entryway,shelf,mantel fireplace,dining table, andkitchen counter. - For Amazon Italy bed ads, discard non-target-language or wrong-product terms unless the user explicitly wants them. Examples from evidence:
toddler bed travelis not a Montessori floor bed term;culla,lettino neonato,materasso,letto contenitore,divano letto,letto a castello, andusatoare negative candidates when not sold.
- For 华伦料理棒,
5. Layer outputs and actions by match type or use case
- Make the output directly usable: separate add/keep/remove, exact/phrase/broad, positive/negative, or core/long-tail/test terms.
- Concrete conventions from evidence:
- For Amazon Italy advertising, use
精准,词组, and广泛as distinct match-type layers; for negatives, separate否定词组from否定精准. - For negative keywords in the Italy bed case, put broad unsuitable categories in
否定词组and reserve否定精准for exact blockers such asikea,materasso, andculla. - For 聚光 keyword optimization, compare planning keywords and actual keywords in three groups: intersection terms already delivered, planning terms not yet delivered, and actual terms absent from planning data.
- For Search Term drafting, combine data-supported terms with exploration terms only after classifying which are front-end duplicates, language variants, Spanish variants such as
jarrones/floreros/centro de mesa, style terms, placement terms, and botanical-use terms.
- For Amazon Italy advertising, use
6. Edit configs or generate importable artifacts safely
- If the task modifies files, use precise replacement and only touch the requested keyword list or specified field.
- Concrete conventions from evidence:
- For the war-news classifier, replacing the war list changed the list from
74to51keywords while AI keywords stayed at79; always validate target-list count and unchanged sibling-list count after editing. - When the user says a list should “仅保留” specific items, the old list is not a merge base: remove unlisted items such as
石油,原油,OPEC,欧佩克,油价,天然气,布伦特, andWTIeven if they are semantically related. - When creating keyword tables for ads, keep bid currency and match type together: examples include
€0.11,€0.23,€0.26,词组,精准, and广泛; do not output a keyword without its intended matching or bidding handling when the user is configuring a campaign. - If exporting a spreadsheet or list, keep one row per keyword unless the user asks for grouped prose; include rule/status columns such as
新增,保留,暂停,否定,测试, or待验证.
- For the war-news classifier, replacing the war list changed the list from
7. Validate quantities, conflicts, and claim boundaries
- Before final delivery, verify counts, formulas, duplicates, conflicts, and whether each conclusion is fact, assumption, or next test.
- Concrete conventions from evidence:
- For config updates, print or summarize the final keyword count, newly added items, removed items, and unaffected keyword groups; mention that historical records are not reclassified unless the system actually backfills them.
- For 聚光 analysis, validate that totals reconcile with source totals such as
总计rows by recomputing消费,展现量,点击量,进店uv, and derivedcpuv; do not average row-levelcpuvto get totalcpuv. - For Search Term choices, check that excluded terms are excluded because of a stated rule: front-end duplicate, poor relevance, off-position style such as
farmhouse rustic vintagefor a modern coastal vase, competitor risk, wrong language, or missing evidence. - For bid or budget recommendations, separate observed facts (
actual cpclower than市场出价,进店uv=0,cpuvundefined) from hypotheses (broad match may be drifting,decision-stage users may convert) and from tests to run.
Boundaries and prohibitions
- Do not produce a final keyword plan before confirming platform, site, purpose, and scope when those are missing; for 聚光, distinguish
聚光数据汇总from聚光定向消耗and honor the stated window such as7.9-8.3. - Do not ignore sheet names, headers, field definitions, date ranges, or user-stated exclusions; if source fields like
搜索主题,消费,展现量,点击量,进店uv,cpuv,月搜索指数, or市场出价are absent, state the limitation instead of calculating substitutes. - Do not invent metrics, search volume, CPC, ACOS, ROAS, SOC/SOV, or conversion data when the source lacks them;
#DIV/0!incpuvorcpcmeans the denominator is zero, not a usable numeric value. - Do not put competitor, off-language, or visibly unrelated terms into core targeting unless the user explicitly requests conquesting or exploration; keep
博朗料理棒as competitor-test, not core brand traffic for华伦料理棒. - Do not update unrelated config logic, sibling keyword groups, schedules, document IDs, or scripts when only keywords are requested; in the war-news case,
美伊战争keywords changed from74to51, while AI keywords remained79. - Do not use summary or pre-aggregated tables as the sole authority when row-level data is available; recompute totals from rows with
消费,点击量, and进店uv, then use总计only to reconcile.
Minimal example
User asks: “这个是聚光平台关键词规划工具的数据,结合我们投的关键词分析,品牌是华伦料理棒。”
Response shape:
- Confirm sources: planning fields
月搜索指数/市场出价/竞争指数and actual fields搜索主题/消费/点击量/进店uv/cpc/cpuv. - Recompute formulas:
cpc=消费/点击量,cpuv=消费/进店uv,转化率=进店uv/点击量. - Split terms into: already delivered intersection terms, planning-only opportunities, actual-only terms.
- Mark actions: add budget to low-
cpuvdirect terms, test competitor/decision terms separately, pause zero-进店uvbroad terms, and list assumptions separately from facts.