Google has quietly fixed one of Discover's oldest complaints: recommendations that are flat-out old news.
A new paper from Google researchers describes SDF (Supersession-Decay Filtering), a system now fully deployed in Google Discover, the personalized feed with hundreds of millions of daily users. SDF targets staleness through two separate mechanisms: supersession, where a newer story makes an older one obsolete, and decay, where a story's relevance simply fades over time. It uses two learned models working together, one that detects when one item supersedes another and one that predicts an item's future traffic from its content, then prunes stale candidates before they ever reach the ranking stage. Over two years in production, user-filed staleness complaints dropped 54.9% compared to before deployment.
This matters because most recommendation systems still lean on crude proxies for freshness: hard age cutoffs or engagement stats that only reveal a story is stale after users have already seen it and moved on. SDF's bet is that content staleness is predictable from the content itself, not just from how people react to it after the fact. That is a meaningfully different approach for any feed-based product, from news aggregators to social platforms, that struggles with the same lagging-signal problem.
Worth noting: these are Google's own numbers from Google's own paper, not an independent audit, so treat the 55% figure as a company-reported result rather than a verified benchmark.