Skip to content

Chapter 3 of 7

When Optimization Goes Wrong

Optimizing ranking for engagement can produce surprising side effects. Proxy gaming occurs when creators or content learn to exploit a measurable proxy, such as watch time, in ways that do not actually improve the underlying goal like satisfaction or wellbeing. A related phenomenon is the satisfaction gap: content that secures clicks or attention but leaves viewers feeling worse, misled, or frustrated, which erodes long-term trust in the platform. To detect and correct for these problems, systems track regret or "I didn't want this" signals, since high engagement can still be harmful if users regret the time spent.

These patterns are examples of metric myopia, where over-optimizing a single metric such as watch time degrades the overall experience, leaving users burned out or surrounded by low-quality content. The same idea is captured by Goodhart's law applied to social algorithms: when a measure becomes a target, it ceases to be a good measure, because creators and gameplay distort the metric itself.

The practical lesson for designers is that alignment between measured signals and real value is fragile. Counterweight signals like regrets, surveys, and long-term retention help platforms balance the short-term appeal of engagement against the long-term cost of dissatisfaction, content fatigue, or trust loss.

All chapters
  1. 1How Feed Recommendation Systems Work
  2. 2Ranking Signals and Their Tradeoffs
  3. 3When Optimization Goes Wrong
  4. 4Creator Growth and Content Design
  5. 5Distribution Surfaces and Platform-Specific Optimization
  6. 6Content Policy, Integrity, and Manipulation
  7. 7Trends, Identity, and Practical Strategies

Drill it

Reading is not remembering. These come from the Social Media Scrolling Virality Addon 60 V2 deck:

Q

SM: What is a “candidate set” in feed ranking?

The shortlist of posts/videos retrieved for you before final ranking decides what gets shown next.

Q

SM: What is “retrieval” in recommendation systems (simple)?

The step that quickly finds potentially relevant items from a huge catalog before more expensive ranking.

Q

SM: Why do platforms use multiple stages (retrieval → ranking)?

To scale: retrieval narrows billions of items to thousands; ranking then scores the shortlist more precisely.

Q

SM: What is a “feature” in ML ranking models?

An input variable used to predict outcomes (e.g., your recent watches, video length, creator relationship).