Designing Effective Ad Mixing in Short‑Video Feed Streams

Short‑video feed streams such as TikTok, Kuaishou, and WeChat Channels deliver immersive, infinite‑scroll content. The feed acts as an attention‑harvesting mechanism by removing selection friction and end signals, thereby maximizing per‑user watch time for ad monetization or e‑commerce conversion.

A straightforward implementation fetches a fixed number of videos (e.g., ten), retrieves a couple of ads from an ad service, and manually inserts the ads at predetermined positions (e.g., third and seventh). This approach leads to duplicate ad exposure when new videos are published between page requests: the first‑page ads reappear on the second page because the underlying data set shifts, causing user annoyance.

To eliminate this duplication, the article proposes using cursor‑based pagination. By advancing a cursor rather than relying on offset indexes, the system can fetch the next page without being affected by newly inserted videos, guaranteeing that each ad appears only once regardless of data churn.

Fetching ads synchronously with video data is also discouraged. Video content must respond within tens of milliseconds, while ad retrieval—dependent on user profiling—can take several hundred milliseconds or even up to a second. Mixing these calls would degrade video latency and harm user experience.

The recommended architecture separates concerns: the backend feed gateway returns only placeholder ad slots, indicating where an ad should appear. The client, upon recognizing an ad slot, invokes the ad SDK to request the actual ad. This client‑side mixing isolates the short‑video service from the ad service, so a failure in either service does not impact the other, and it ensures that billing occurs only after the client confirms effective ad exposure.