返回提示词目录第 1308 页

设计广告地域留出增量测试

依据区域销售与投放数据,制定实验分组、周期和分析方案,评估广告带来的增量销售。

完整提示词

<context>
The user wants to know whether an ad channel actually causes extra sales, not just whether the platform claims them. Platform attribution takes credit for people who would have bought anyway, especially in retargeting, brand search and broad automated campaigns. A geo test turns ads off (or on, or up) in some regions and compares them with similar regions that keep the normal plan. What experts know: regions must be matched on pre-period sales trends, not just size; the outcome must be measured in the business's own sales data by region, not platform conversions; the test needs to be long enough to cover purchase lag and enough regions or volume to detect a realistic effect; spillover (people crossing regions, national TV, delivery addresses differing from where ads were seen) blurs results; other changes during the test (promotions, price changes) must be the same in both groups; and the readout is incremental sales per unit of spend, compared with what the platform reported. Open-source tools for synthetic control and matched markets exist and can help, but the design choices matter more than the tool.
</context>

<task>
Question: {{question}}
Business data:
<business_data>
{{business_data}}
</business_data>
Spend: {{spend}}
{{#constraints}}Constraints: {{constraints}}{{/constraints}}

1. Question and metric: restate the question as a testable one, choose the primary outcome (orders, revenue, new customers, leads from the business's own system) and secondary metrics, and the hypothesis with the effect size that would change a decision.
2. Region selection: the unit (state, city, postcode area, store catchment), how to group into test and control by matching pre-period trends and size, regions to exclude (too small, too unusual, home market), and a pre-period check of how closely control tracks test.
3. Test design: holdout (turn off), lift (turn on or increase), or scaled spend levels; what changes in test regions and what stays the same everywhere; how to apply geo targeting in the platform and check it holds.
4. Size and duration: pre-period length, test length covering purchase lag plus a cool-down, and a rough power reasoning: given weekly sales noise in the history, what effect size the design can detect. If the data cannot detect a realistic effect, say so and suggest a bigger cut, more regions or a longer test.
5. Guardrails: spillover controls, freeze on other changes, a stop rule if business harm exceeds a limit, and a log of anything unusual.
6. Readout plan: the method (difference in differences or a synthetic control, with a plain explanation), confidence interval, incremental sales and incremental cost per sale or incremental ROAS, comparison with platform-reported results, and the decisions each outcome leads to.
</task>

<constraints>
- Do not invent sales figures, effect sizes or variance; use the user's data or ask for weekly regional history. Mark rough power estimates as approximate.
- Do not claim a result is significant without the calculation; describe uncertainty honestly.
- You may name open-source geo testing tools, without links; do not claim features you are not sure of.
- If sales cannot be tied to regions, say a geo test is not feasible and suggest an alternative (platform conversion lift study, time-based on and off test with caveats).
</constraints>

<output_format>
## Question and metric
Hypothesis, primary metric, decision threshold.

## Region selection
Unit, grouping method, a table: Group | Regions | Share of sales | Pre-period fit.

## Test design
What changes where, and how to set it up in the platform.

## Size and duration
Timeline (pre-period, test, cool-down) and the power reasoning.

## Guardrails
Checklist.

## Readout plan
Method, outputs, and decisions by outcome.
</output_format>

可替换内容

  • {{question}}
  • {{business_data}}
  • {{spend}}
  • {{#constraints}}
  • {{constraints}}
  • {{/constraints}}

分类与标签

商业营销、广告效果评估、地域实验、增量销售、投放归因

前往 AI 图片创作