千川人群定向拓展

要解决的具体问题

千川投放中的人群扩量常被简化成两个数字：人群包有多大，最近三天ROI是否达标。真正落地时会出现三类故障：

• 高频次低效人群持续抢预算，曝光很多但成交不成比例；
• 高ROI来自归因延迟或新客偶然成交，把偶然当成稳定结构；
• 不同账户、商品、归因窗口、人群包版本混在一起评分，决策口径失真。

这些做法会让定向越调越乱，团队复述不出“为什么纳入这一批人”。

技能如何工作

该技能把扩量从拍脑袋变成可审计的分层决策。首先对齐账户、商品、归因窗口、人群包版本和新客定义，避免不同口径直接合并。接着收集人群规模、触达、曝光、消耗、成交、订单与新客订单，并确认目标ROI和最低样本要求。

随后运行 scripts/plan_audience_expansion.py，生成人群分层和默认 dry_run 的定向动作清单。对高频低效人群优先降频或排除，只有高ROI且高新客占比的人群才会进入拓展候选。输出 audience_decisions.csv 记录人群ROI、新客率、频次、分层、建议动作和证据；输出 action_manifest.json 保留纳入、排除、保留或复测动作，默认 dry_run: true。字段和阈值说明见 references/method.md。

适用边界与注意点

• 不把相关性描述为因果；扩量前要求可比较的测试窗口。
• 不把低消耗高ROI直接视为优胜人群，必须结合频次、新客率与样本量复核。
• 明确列出样本不足、归因延迟、重叠人群和人群包授权风险。
• 真实更新人群定向前必须取得授权、校验人群包ID并核验平台回执；草案不等于已生效。
• 不读取或输出 Access-Token、Cookie、密码或密钥。

该技能适合需要解释决策依据、控制扩量风险的千川投放复盘与定向治理场景。

The Specific Problem It Addresses

Audience expansion in Qianchuan campaigns is often reduced to two numbers: how large an audience pack is, and whether ROI in the last three days clears a threshold. In practice, three failure modes appear:

• High-frequency, low-efficiency audiences keep taking budget, generating impressions without proportional conversions.
• A high ROI number comes from attribution delay or accidental new-customer orders, turning chance into a false stable pattern.
• Different accounts, products, attribution windows, audience-pack versions, and new-customer definitions are scored together, corrupting the decision baseline.

These habits make targeting messier over time, and teams cannot explain why a specific group was included.

How the Skill Works

The skill turns expansion from guesswork into an auditable layered decision process. It first aligns account, product, attribution window, audience-pack version, and new-customer definition so mixed calibers are not merged into one score. It then collects audience size, reach, impressions, spend, conversions, orders, and new-customer orders, and confirms the target ROI and minimum sample requirement.

Next, it runs scripts/plan_audience_expansion.py to generate audience tiers and a default dry_run action manifest. High-frequency, low-efficiency audiences are prioritized for frequency reduction or exclusion; only high-ROI audiences with a strong new-customer share become expansion candidates. audience_decisions.csv records audience ROI, new-customer rate, frequency, tier, suggested action, and evidence. action_manifest.json keeps include, exclude, retain, or retest actions, with dry_run: true by default. Field and threshold details are documented in references/method.md.

Applicable Boundaries and Considerations

• Correlation is not treated as causation; expansion requires comparable test windows before action.
• Low spend with high ROI is not automatically a winning audience; frequency, new-customer share, and sample size must be reviewed together.
• Insufficient samples, attribution lag, overlapping audiences, and audience-pack authorization risks are listed explicitly.
• Before a real targeting update, authorization must be obtained, audience-pack IDs must be validated, and platform receipts must be checked; a draft is not a live change.
• Access tokens, cookies, passwords, and keys are neither read nor output.

The skill fits Qianchuan campaign reviews and targeting governance where teams need to justify decisions and control expansion risk.