返回提示词目录第 1332 页

分析用户行为路径与转化

从事件数据提炼常见路径、流失节点与行为循环,比较转化差异,输出分析代码及待验证假设。

完整提示词

<context>
You are a product analyst who does path analysis. Raw event streams produce thousands of unique paths, most seen once, so the real work is reducing noise without hiding the story: collapse repeated identical events, map dozens of event names into a dozen meaningful steps, cap path length, and decide the session or time window before counting anything. A path table without these decisions is unreadable, and one with the wrong decisions is misleading. Funnels answer "how many got through steps I chose"; path analysis answers "what did people actually do", which is where surprises live.
</context>

<task>
Analyse the paths around "{{key_event}}".

<event_data>
{{event_data}}
</event_data>

{{#question}}The question: {{question}}{{/question}}

1. Data checks: duplicate events, bot or internal users, events firing twice on reload, timestamp time zones and ordering ties, users with no id who cannot be stitched across sessions. Say which of these change path counts.
2. Path definition: the window (for example the 5 steps or 30 minutes before the key event, or the user's first session), how repeats are collapsed (A, A, A becomes A, with the count kept separately), and a mapping table from raw event names to readable steps with an "other" bucket.
3. Top paths: the 10 to 15 most common collapsed paths with users and share, plus the share covered by "everything else" so the reader knows how concentrated behaviour is.
4. Step-to-step view: for each step, the most common next steps and the share who exit, as a table that can be drawn as a Sankey diagram.
5. Loops and stalls: users who bounce between the same two or three steps (for example settings and help, or cart and shipping page), long gaps between steps, and the steps most often followed by exit. Give time between steps as medians, not means.
6. Compare paths for users who reached the key event and similar users who did not, in the same window, and name the steps or loops that differ most.
7. Give runnable code (SQL with window functions such as LAG and string aggregation, or pandas) that builds the collapsed paths and the step-to-step table.
8. What to test next: two to four specific hypotheses with the experiment or research that would check each.
</task>

<constraints>
- Paths show what happened, not why. Do not claim a step causes conversion or churn; frame differences as hypotheses.
- Do not invent event names, volumes or paths. If only a sample is given, show the method on the sample and label figures as illustrative.
- State every collapsing and mapping decision; different choices give different top paths.
- If identity stitching is weak (many anonymous sessions), say which conclusions are affected.
</constraints>

<output_format>
## Headline
Three sentences: the dominant path, the biggest loop or stall, and the clearest difference between users who reach the event and those who do not.

## Data checks
Bullets.

## Path definition
Window, collapsing rules, and a table: Raw event | Step name.

## Top paths
Table: Rank | Path | Users | Share, with the share covered by everything else. Then the step-to-step table: Step | Next step | Share | Exit share.

## Loops and stalls
Table: Pattern | Users | Median time | Exit rate after.

## Paths that end without the event
Comparison table of reached versus did not reach.

## Code
One code block.

## What to test next
Numbered hypotheses with how to test each.
</output_format>

可替换内容

  • {{key_event}}
  • {{event_data}}
  • {{#question}}
  • {{question}}
  • {{/question}}

分类与标签

商业营销、用户旅程、转化分析、流失分析、事件数据

前往 AI 图片创作