Date Written: August 17, 2023
Keywords: Influencer Advertising, Video Advertising, Entertainment Commerce, Creator Economy, Sales Conversion, Incentive Alignment, Computer Vision, TikTok
Summary and Analysis of “Engagement that Sells: Influencer Video Advertising on TikTok”
By Jeremy Yang (Harvard), Juanjuan Zhang (MIT), Yuhan Zhang (BTBU)
🧠 Core Insight:
Not all engagement is good engagement—especially in influencer video ads. What matters most for actual sales lift is whether the product itself is what’s engaging viewers. This paper introduces a novel, scalable algorithm to calculate a Product Engagement Score (PE-score) that predicts which influencer TikTok videos will actually drive purchases—not just rack up likes.
🔍 Key Concepts:
- The Problem: Misaligned Incentives
- Influencers are incentivized to promote themselves, not the product.
- High engagement (likes, shares) ≠ high conversion.
- Sellers lack tools to assess if a video’s engagement is product-relevant.
- The Solution: Product Engagement Score (PE-score)
A score that reflects how engaging the product itself is within the video, computed in 3 steps:
- Engagement heatmap: Pixel-level saliency using a deep 3D convolutional neural network (CNN) trained on engagement data (e.g., shares).
- Product heatmap: Object detection (via SIFT) marks where/when product appears in video.
- PE-score: Normalized dot product of both maps = engagement-weighted presence of the product.
📊 Key Findings:
🧪 Validation:
- Dataset: 2,685 influencer video ads on TikTok, sales tracked via Taobao.
- Result: PE-score significantly predicts sales lift.
- Engagement alone: No correlation with sales.
- Product placement alone: No correlation.
- PE-score: Strong correlation, even controlling for influencer popularity, price, etc.
- Influencers advertising their own products scored 31% higher PE-scores on average → aligns incentives = higher effectiveness.
🧠 Behavioral Insights:
- Peak viewer attention: First 6 seconds.
- Product + human face = higher attention.
- Facial expressions (happy, sad), novelty, and dynamic motion boost saliency.
💡 Practical Applications:
For Brands:
- Screen influencer content pre-release using PE-score to predict conversion.
- Write contracts based on PE-score thresholds, not just engagement.
For Influencers:
- Use PE-score feedback to optimize ad content before publishing.
- Better negotiate fees with proof of performance potential.
For Platforms:
- Improve ad attribution models.
- Power recommendation systems that favor truly effective content.
🤖 Technical Highlights:
- Uses transfer learning with Xception CNN architecture.
- Analyzes videos frame-by-frame and pixel-by-pixel (after reducing size/FPS).
- Model trained on >16,000 TikTok ads, tested on a separate panel with real sales data.
- All computations can be done before ad release, making it scalable for real-world use.
🔁 Strategic Implications:
This paper helps shift performance marketing from proxy metrics (likes, shares) to true outcome metrics (sales lift). It’s a call for alignment between content creation and commerce, especially as platforms like TikTok blur the line between entertainment and shopping.
🧭 TL;DR (for Decision Makers):
