方法与参与度 / Methodology
先确认样本怎么来，再解释画像可信度。
原报告刻意平衡抽取 Power、Regular、Casual 三层用户，但总体占比需要重新加权。这个页面集中放样本构成、参与度和低信号风险。
Caveats
四个解释前提
01 · 样本只代表主动触发 Conversate AI cue 的用户，不代表全部眼镜客户。The sample represents Conversate-engaged users, not the entire smart-glasses customer base.
02 · 转录里有多位说话者，无法确定佩戴者身份；画像来自跨多次会话的重复语境。Transcripts contain multiple speakers; personas are inferred from repeated contexts, not wearer-attributed utterances.
03 · 每个 tier 只有 70 个样本，低于 5 人的细分计数只能作方向性参考。Each tier has 70 sampled users; small counts below five are directional rather than statistically reliable.
04 · Power 用户被刻意过采样；总体占比使用真实 tier mix 重新加权后才适合对外引用。Power users are intentionally oversampled; user-base shares are reweighted to the true tier mix.
Sample Composition
样本与总体
| 层级 / Tier | 符合条件用户 / Eligible | 总体占比 / Share | 样本数 / Sampled | 抽样率 / Rate |
|---|---|---|---|---|
| Power | 1,444 | 12.7 % | 70 | 4.85 % |
| Regular | 5,370 | 47.2 % | 70 | 1.30 % |
| Casual | 4,562 | 40.1 % | 70 | 1.53 % |
Evidence Rule
3-session 证据门槛
每个用户的完整会话历史先被汇总，再提取至少跨 3 次会话重复出现的主题与使用场景。相似标签被合并为统一的 canonical set，用于后续 rollup。
Each sampled user's session history was summarized, then recurring themes and use cases were consolidated into canonical labels.
Engagement by Tier
参与度分层
| 层级 / Tier | 用户数 / Users | 会话 p25/med/p75 | 活跃天数中位数 | 消息数中位数 | 单会话消息中位数 | Cue 展示中位数 | Cue 打开率 p25/med/p75 |
|---|---|---|---|---|---|---|---|
| Power | 70 | 33 / 39 / 60 | 15 | 13,315 | 284 | 736 | 2% / 66% / 97% |
| Regular | 70 | 7 / 11 / 15 | 5 | 1,612 | 132 | 106 | 1% / 51% / 97% |
| Casual | 70 | 2 / 3 / 3 | 2 | 117 | 39 | 12 | 0% / 5% / 99% |
Low Signal
无法形成画像信号的比例
| 层级 / Tier | 低信号用户 / Low signal | 层内占比 / Tier share |
|---|---|---|
| Power | 1 / 70 | 1% (1) |
| Regular | 7 / 70 | 10% (7) |
| Casual | 54 / 70 | 77% (54) |
Risk
Casual 层最需要谨慎
Casual 用户只有 2 到 4 次会话，很多样本没有足够重复语境可判定画像。Casual 相关的 persona 结论应当弱化使用。
Casual-tier users often lack enough repeated context, so persona claims for this tier carry the highest uncertainty.