🎆 Meta explains how it’s improving Reels recommendations

Meta has shared details on how it’s refining Reels recommendations by adding large-scale user surveys into its ranking system, instead of relying only on likes, shares, and watch time.

The idea is simple. Ask users directly how a Reel made them feel, then feed that data into the model alongside traditional engagement signals.

What changed

✔️ In-feed surveys are shown between Reels to capture real-time user sentiment
✔️ Survey data is adjusted for bias and nonresponse to better reflect actual preferences
✔️ Recommendations now use explicit feedback, not just passive behavior

Results so far
➡️ Alignment with true user interests improved from about 48% to over 70%, according to Meta
➡️ Reels are reportedly becoming more personalized and better at driving repeat visits

Meta admits there is still work to do, especially for users with limited engagement history and for improving content diversity.

TikTok remains the benchmark, largely due to deeper content understanding through visual and entity recognition inside videos. Meta’s approach is more conservative, leaning on surveys and standard signals rather than deep computer vision.

If Reels have felt more relevant recently, this shift is likely the reason.

✅ Follow @inst