Six Google researchers describe SDF in a scientific paper: the filter that, they say, Google Discover has used for two years to remove stale items from the feed before ranking it. It matters to any site that gets Discover traffic: a piece superseded by newer coverage, or one that has run out of interest, stops being eligible.
What we know
- The paper: "Decomposing Staleness in Recommender Systems", by six authors at Google LLC, posted to arXiv on August 16, 2026 and accepted at CIKM '26, held in Rome on November 7-11.
- Where it sits: per the paper, SDF is "fully deployed" in Discover and prunes candidates before the ranking stage. An item flagged by either filter does not get ranked.
- Filter 1, supersession: it compares each new item with existing ones about the same event or entity. It only counts when the new item contradicts, supersedes or shifts the narrative; another piece on the same topic is not enough. A language model writes its reasoning first and then the label.
- Filter 2, relevance decay: it predicts from the content how much traffic an item has left, trained on items' lifetime visits. The model predicts at several horizons, from 12 hours to 14 days: breaking news is measured in hours and slower-decaying pieces out to two weeks. Evergreen items, such as a health explainer, are not filtered.
- Not personalized: filter outputs are the same for every user and are computed off the request path.
- The numbers: user reports of stale content fell 54.9% against the pre-launch baseline (64.0% for superseded items, 34.4% for decayed ones). In a two-month holdback, positive engagement rose 0.26%, dismissals fell 0.16% and feed diversity rose 0.13%. The two filters save 4.76% and 3.98% of compute hours.
What changes and what doesn't
The paper describes how Discover works; it is not Google guidance for publishers and not an announced change. Every figure comes from the authors about their own system, with no outside measurement, and the paper says nothing about Search results. Search Engine Journal covered it on October 21. That Discover favors recent content is not new; what is new is that there are two separate reasons to drop out, and that evergreen content sits outside the decay filter.
How to tell if it affects you
- Check whether you get Discover traffic. In Search Console, the Discover report only appears once a site reaches a minimum number of impressions. If it isn't there, this filter changes nothing for you today.
- Measure each piece's lifespan. Filter the report by page and look at daily impressions after publication. A news item that drops to zero within a day or a few days fits what the paper describes; it is not a penalty.
- Separate what expires from what doesn't. A deal, an event or a breaking update is short-lived by design. Guides and explainers that stay useful months later are, per the paper, the kind the decay filter leaves alone. That mapping is this article's inference: the paper gives no rules for publishers.
- Watch fast-moving topics. If you cover something that is still developing, a later piece that changes the story can push yours out. The paper does not say whether updating the same page brings it back.
- Don't refresh dates without refreshing content. The decay filter reads the content, not just the date, according to the paper.
Related: Search Now Brings Just 12% of Publisher Pageviews
Sources
- Di Bai et al. (Google), "Decomposing Staleness in Recommender Systems: A Dual-Filter Framework for Supersession and Decay", arXiv 2608.15780, August 16, 2026
- Search Engine Journal, Roger Montti, October 21, 2026
Updates: if Google documents SDF in its Discover guidance, it will be added here with a link.