algo-ad-bidding
挑選廣告出價策略，從手動 CPC 到 target-CPA、ROAS
中文翻譯
---
name: "algo-ad-bidding"
description: "實作並挑選廣告出價策略，從手動 CPC 到自動化的 target-CPA、target-ROAS 都包含。當使用者需要選出價策略、設定自動出價或最佳化出價參數時使用，即使他們說的是『what bidding strategy should I use』『target CPA setup』或『smart bidding configuration』。"
metadata:
category: "WP-37 廣告演算法"
tags: ["advertising", "bidding-strategy", "automated-bidding", "optimization"]
---
# 廣告出價策略
## 概述
出價策略決定廣告主每次競價要付多少錢。範圍從手動 CPC（完全自己控制）到自動化策略（Target CPA、Target ROAS、Maximize Conversions），後者用機器學習依情境訊號即時最佳化出價。
## 何時使用
**觸發條件：**
- 在手動與自動出價策略之間做選擇
- 設定或排查 Target CPA / Target ROAS 廣告活動
- 分析出價策略的成效並做調整
**不該使用的情況：**
- 設計競價機制本身時（用 GSP/VCG）
- 建立 CTR 預測模型時（用 CTR prediction skill）
## 演算法
```
IRON LAW: Automated Bidding Requires SUFFICIENT Conversion Data
Below ~30 conversions/month, the algorithm lacks signal and performs
WORSE than manual bidding. Strategy selection depends on data volume:
- < 30 conv/month: Manual CPC or Maximize Clicks
- 30-50 conv/month: Maximize Conversions
- 50+ conv/month: Target CPA
- 50+ conv/month + revenue data: Target ROAS
```
### 階段 1：輸入驗證
評估：每月轉換量、轉換追蹤的準確度、廣告活動預算、業務目標（量 vs 效率 vs 營收）。
**關卡：** 轉換追蹤已驗證，且資料量足以支撐所選策略。
### 階段 2：核心演算法
**Manual CPC：** 逐個關鍵字設定出價。依裝置、時段、地點、受眾成效資料來調整。
**Target CPA：** 1. 設定目標單次獲客成本。2. 演算法用情境訊號預測每次競價的轉換機率。3. 轉換機率高就加價，低就降價。4. 目標是長期平均下來落在 target CPA。
**Target ROAS：** 跟 CPA 一樣，但改成最佳化廣告投資報酬率 = conversion_value / cost。
### 階段 3：驗證
監控：實際 CPA 對比目標、轉換量是否穩定、曝光比重變化、預算使用率。
**關卡：** 學習期（2-4 週）過後，實際 CPA 要落在目標的 20% 以內。
### 階段 4：輸出
回傳策略建議，附上預期的成效範圍。
## 輸出格式
```json
{
"recommendation": {"strategy": "target_cpa", "target": 500, "currency": "TWD", "confidence": "high"},
"expected_performance": {"cpa_range": [400, 600], "volume_change": "-10% to +15%"},
"metadata": {"monthly_conversions": 85, "current_cpa": 550, "learning_period_days": 14}
}
```
## 範例
### 輸入輸出範例
**輸入：** 電商廣告活動，每月 120 次轉換，目前 CPA=NT$450，目標：維持 CPA、拉高量
**預期：** Target CPA 設在 NT$450。預期：演算法找到有效率的競價後，量會增加 10-20%。
### 邊界情況
| 輸入 | 預期 | 原因 |
|-------|----------|-----|
| 每月 10 次轉換 | Manual CPC | 資料不足，無法自動化 |
| Target CPA 設得太激進 | 量掉到接近零 | 演算法找不到有利潤的競價 |
| 轉換追蹤壞掉 | 所有策略都會失敗 | 垃圾資料進，垃圾最佳化出 |
## 注意事項
- **學習期的波動**：切換策略後的前 2 週成效不穩定。這段時間不要改目標。
- **轉換延遲**：如果轉換要隔好幾天才歸因（例如 B2B），演算法就是在過期的資料上做最佳化。用轉換建模，或把歸因視窗拉長。
- **預算是限制條件**：Target CPA 達不到目標就不會花錢。大預算搭配很激進的 CPA，不會讓花費變多，只是省錢。
- **微轉換**：如果拿代理轉換（加入購物車）而不是最終購買來訓練，演算法就會去最佳化那個代理指標。要確保追蹤的轉換和商業價值一致。
- **季節性衝擊**：自動出價是從近期資料學習的。黑色星期五、節日或競爭事件會把它帶偏。要用季節性調整。
## 參考資料
- 出價策略遷移的操作手冊，見 `references/migration-playbook.md`
- 學習期的最佳實務，見 `references/learning-period.md`英文原文
---
name: "algo-ad-bidding"
description: "Implement and select ad bidding strategies from manual CPC to automated target-CPA and target-ROAS. Use this skill when the user needs to choose a bidding strategy, set up automated bidding, or optimize bid parameters — even if they say 'what bidding strategy should I use', 'target CPA setup', or 'smart bidding configuration'."
metadata:
category: "WP-37 廣告演算法"
tags: ["advertising", "bidding-strategy", "automated-bidding", "optimization"]
---
# Ad Bidding Strategies
## Overview
Bidding strategies determine how much an advertiser pays per auction. Range from manual CPC (full control) to automated strategies (Target CPA, Target ROAS, Maximize Conversions) that use ML to optimize bids in real-time based on contextual signals.
## When to Use
**Trigger conditions:**
- Choosing between manual and automated bidding strategies
- Setting up or troubleshooting Target CPA / Target ROAS campaigns
- Analyzing bid strategy performance and making adjustments
**When NOT to use:**
- When designing the auction mechanism itself (use GSP/VCG)
- When building a CTR prediction model (use CTR prediction skill)
## Algorithm
```
IRON LAW: Automated Bidding Requires SUFFICIENT Conversion Data
Below ~30 conversions/month, the algorithm lacks signal and performs
WORSE than manual bidding. Strategy selection depends on data volume:
- < 30 conv/month: Manual CPC or Maximize Clicks
- 30-50 conv/month: Maximize Conversions
- 50+ conv/month: Target CPA
- 50+ conv/month + revenue data: Target ROAS
```
### Phase 1: Input Validation
Assess: monthly conversion volume, conversion tracking accuracy, campaign budget, business goal (volume vs efficiency vs revenue).
**Gate:** Conversion tracking verified, sufficient data for chosen strategy.
### Phase 2: Core Algorithm
**Manual CPC:** Set bid per keyword. Adjust based on: device, time, location, audience performance data.
**Target CPA:** 1. Set target cost-per-acquisition. 2. Algorithm predicts conversion probability per auction using contextual signals. 3. Bids up for high-probability conversions, down for low. 4. Aims to average at target CPA over time.
**Target ROAS:** Same as CPA but optimizes for return on ad spend = conversion_value / cost.
### Phase 3: Verification
Monitor: actual CPA vs target, conversion volume stability, impression share changes, budget utilization.
**Gate:** Actual CPA within 20% of target after learning period (2-4 weeks).
### Phase 4: Output
Return strategy recommendation with expected performance ranges.
## Output Format
```json
{
"recommendation": {"strategy": "target_cpa", "target": 500, "currency": "TWD", "confidence": "high"},
"expected_performance": {"cpa_range": [400, 600], "volume_change": "-10% to +15%"},
"metadata": {"monthly_conversions": 85, "current_cpa": 550, "learning_period_days": 14}
}
```
## Examples
### Sample I/O
**Input:** E-commerce campaign, 120 conversions/month, current CPA=NT$450, goal: maintain CPA, increase volume
**Expected:** Target CPA at NT$450. Expected: volume +10-20% as algorithm finds efficient auctions.
### Edge Cases
| Input | Expected | Why |
|-------|----------|-----|
| 10 conversions/month | Manual CPC | Insufficient data for automation |
| Target CPA too aggressive | Volume drops to near zero | Algorithm can't find profitable auctions |
| Conversion tracking broken | All strategies fail | Garbage data → garbage optimization |
## Gotchas
- **Learning period volatility**: First 2 weeks after switching strategies show unstable performance. Don't change targets during this period.
- **Conversion delay**: If conversions take days to attribute (e.g., B2B), the algorithm optimizes on stale data. Use conversion modeling or extend the attribution window.
- **Budget as a constraint**: Target CPA won't spend if it can't hit the target. Setting an aggressive CPA with a large budget doesn't increase spend — it just saves money.
- **Micro-conversions**: If training on a proxy conversion (add to cart) instead of final purchase, the algorithm optimizes for the proxy. Ensure the tracked conversion aligns with business value.
- **Seasonality shocks**: Automated bidding learns from recent data. Black Friday, holidays, or competitive events can throw it off. Use seasonality adjustments.
## References
- For bid strategy migration playbook, see `references/migration-playbook.md`
- For learning period best practices, see `references/learning-period.md`