ads-test

AgriciDaniel ｜ 行銷與 SEO ｜ 原文 ｜ MIT

設計並判讀付費廣告的 A/B 實驗，講清楚假設、樣本數和停止規則

---
name: ads-test
description: "設計並評估付費廣告實驗，包含假設、隨機化單位、樣本數與期間的假設、護欄、平台實驗工具、分析和決策規則。適用於 A/B 測試、split test、實驗設計、假設、統計顯著性、樣本數、測試期間或實驗結果判讀。"
---

# 付費媒體實驗

1. 說明要做的決策、因果假設、實驗組、對照組、隨機化單位、母體、主要指標、護欄、最小效果和停止規則。
2. 檢查平台限制、重疊的實驗、轉換延遲、季節性、互相干擾和衡量品質。
3. 根據宣告的假設計算樣本數和期間；揭露近似的部分。
4. 一次只改一個決策面，除非設計明確要估計交互作用。
5. 事先登記排除條件、品質檢查、分析方式和決策門檻。
6. 判讀結果時，先驗證分組的完整性和資料的完整度，再估計效果和不確定性。
7. 用有版本的 JSON 回傳設定或結果判讀，並附上白話的決策。

不要反覆偷看、一看到有利的結果就停；不要把檢定力不足的雜訊說成贏家；也不要把結論推廣到沒測過的母體。

---
name: ads-test
description: "Design and evaluate paid-ad experiments with hypotheses, randomization units, sample-size and duration assumptions, guardrails, platform experiment tools, analysis, and decision rules. Use for A/B test, split test, experiment design, hypothesis, statistical significance, sample size, test duration, or experiment readout."
---

# Paid Media Experiment

1. State the decision, causal hypothesis, treatment, control, randomization unit,
   population, primary metric, guardrails, minimum effect, and stopping rule.
2. Check platform constraints, overlapping experiments, conversion lag, seasonality,
   interference, and measurement quality.
3. Calculate sample and duration from declared assumptions; disclose approximations.
4. Change one decision surface unless the design explicitly estimates interactions.
5. Pre-register exclusions, quality checks, analysis, and decision thresholds.
6. For readout, verify assignment integrity and data completeness before estimating
   effect and uncertainty.
7. Return setup or readout in versioned JSON with a plain-language decision.

Do not repeatedly peek and stop on a favorable result, call underpowered noise a
winner, or generalize beyond the tested population.