Google has described a Discover filtering system that removes outdated articles before any ranking takes place, according to Search Engine Journal. The system, Supersession-Decay Filtering (SDF), appears in a Google research paper titled “Decomposing Staleness in Recommender Systems: A Dual-Filter Framework for Supersession and Decay.” For publishers, the consequence is blunt: a page that fails the check never competes.
The paper states that SDF is deployed in Discover, the feed of recommended web pages tailored to each user inside the Google Search app. It places the audience at hundreds of millions of daily users and billions of monthly users. Those figures come from Google’s own authors.
SDF works as a gate in front of the ranking stage. The paper says the ranking systems behind it still weigh user engagement and trending topics. Filtering shrinks the candidate pool first, which means the ranker sees fewer items and spends less computation on them.
Two filters do the work. The first targets supersession, where newer information has replaced an older item. The second targets relevance decay, where usefulness fades over time even though nothing newer has appeared. Search Engine Journal reports that failing either filter makes a page ineligible for Discover ranking. Passing both keeps it in the pool.
The paper also explains why Google did not rely on simpler signals. Publication date alone fails because it records when something was posted, not whether it still helps anyone. A hard date cutoff can also block evergreen material that remains useful. Engagement metrics fail for a different reason: they are reactive, and users have usually decided an item is stale well before the data shows it.
The reported results are specific. SDF cut stale recommendations compared with traditional methods, produced modest gains in engagement and content diversity, and lowered computing costs. Across the two years the system has been live, complaints from users about outdated items fell by 54.9 percent. All of these outcomes are Google’s own measurements. The paper offers no independent audit and no breakdown by publisher type, so nobody outside Google can say which sites lost eligibility.
Now separate the paper from the commentary. The argument that the findings should alarm the SEO community comes from Search Engine Journal’s Roger Montti. He also argues that the findings carry implications for content planning and interlinking, and he asks publishers to decide what share of their output is evergreen, what share needs scheduled updates, and what share is short-lived news. Those questions are his interpretation, not statements in the paper.
His interlinking point deserves attention from technical teams. Many link-suggestion plugins and AI tools match pages on semantic similarity. Montti argues that SDF adds freshness as a second dimension to that calculation. That is his reading, and the paper does not address linking directly.
Our analysis, not the paper’s claim: Discover traffic is notoriously uneven, and a binary eligibility gate would explain some of the sudden drops that publishers report without a clear cause. The paper does not claim that, so treat it as a hypothesis worth testing against your own Discover data in Search Console.
A practical audit follows from the mechanism. Sort pages by the type of staleness they risk. Superseded items, such as coverage of a feature that has since been replaced, need a visible update or a pointer to the newer piece. Slowly decaying items need a review schedule. Evergreen pages should state clearly what they cover so a system judging relevance has no reason to mark them as obsolete.
For the next ninety days, teams that depend on Discover should tag every indexed article as evergreen, updatable, or news, then compare Discover impressions across those groups. If the updatable group underperforms, a refresh cycle is the first lever to pull.
Based on reporting by Roger Montti in Search Engine Journal, published October 21, 2026, on a Google research paper about Discover’s Supersession-Decay Filtering system.
