Google has published research describing a freshness system that has been running inside Google Discover for two years. The system, called Supersession-Decay Filtering, or SDF, sits in front of Discover’s ranking stage and removes content it judges to be stale before that content can compete for a place in the feed, Search Engine Journal reported on October 21, drawing on the researchers’ own paper.
Discover is the personalised article feed in the Google app and on Android phones, serving hundreds of millions of daily users and billions of monthly users, according to the paper. What the research makes clear is that a slice of Discover visibility is decided before ranking begins: if either of SDF’s two filters marks a page as stale, the page is ineligible for the feed no matter how well it might otherwise have performed.
How The Two SDF Filters Decide What Is Stale
The first filter handles supersession. It uses AI to compare older and newer content and judge whether newer information has replaced the older page, making the earlier version obsolete. The second filter handles relevance decay: it predicts when a page will naturally lose usefulness over time, even when nothing newer has superseded it. The two work as one system, and a fail on either side removes the page from the candidate set that Discover’s engagement-based, trend-aware ranking systems then order.
The researchers argue that the older ways of spotting stale content were not good enough. Publication-date cutoffs cannot tell whether a page is still useful: a story can be overtaken by events within hours, while evergreen material can stay relevant for years and still be blocked by an arbitrary date limit. Engagement signals have the opposite problem — they are reactive, registering that readers have gone cold on a topic only after users have already encountered the stale recommendations.
What Two Years Of SDF Did To Discover
According to the paper, SDF reduced stale content recommendations compared with the traditional methods it replaced, and user engagement and content diversity showed modest gains. Over the two years the system has been deployed, user complaints about stale content fell by 54.9%. Pruning outdated pages before the ranking stage also cut computing costs, because the heavier ranking computation runs on a smaller, fresher candidate set.
The paper is titled “Decomposing Staleness in Recommender Systems: A Dual-Filter Framework for Supersession and Decay” and is published on arXiv. Search Engine Journal’s reading of it stresses that the system is not a proposal or a test: it is described by its own authors as deployed in production, which means publishers have been competing under it without knowing its shape.
What Publishers Should Take From The Paper
For publishers, the practical questions the paper raises are about the mix of a site’s output. Evergreen content sustains traffic on queries that stay useful, time-limited content needs a plan for updating it when it goes stale, and news has a short window in which it must be covered while it is still the current version of events. Because SDF judges whether newer material has superseded an older page, an out-of-date article that a publisher never updates is not merely weaker in the feed — under this system it can be filtered out entirely once a fresher version exists.
The freshness dimension also reaches into internal linking. Many linking systems, human and automated, work on semantic relevance alone; the paper suggests Discover’s pipeline adds a time dimension to that calculation, so linking readers toward a superseded page points them at content the feed itself may already have retired. The disclosure lands in the same month as other windows into how Google’s AI surfaces treat sources, including the source preferences button Google is testing inside AI Mode citations and the change that moved links higher in AI Overviews and AI Mode.
Why It Matters
Google Discover’s SDF system matters because it moves freshness from a ranking nudge to a gate. Publishers who assumed an older article could still surface when engagement spiked now face a filter that can remove that article before engagement is ever measured, and the deciding comparison is against the newest version of the story anywhere in the feed, not against the publisher’s own archive. Sites that update stories as facts change, retire or refresh pages whose usefulness has decayed, and keep their evergreen stock genuinely current are aligned with how the system is described; sites that leave superseded pages untouched are, on the paper’s account, quietly losing eligibility. The disclosure also gives the industry a rare named mechanism behind the volatility publishers report after events move on, sitting alongside confirmed ranking changes such as Google’s September 2026 spam update as part of the machinery that decides what readers see.
Background
Google Discover grew out of the Google Feed launched in 2017 and was renamed Discover in 2018, becoming one of the largest referral sources for news publishers as search itself shifted toward AI-generated answers. Google rarely publishes the internal mechanics of the feed, which is why a peer-style research paper describing a production system — complete with its two-year deployment history and complaint figures — is an unusual level of detail for publishers and SEOs to work from.
Source: Search Engine Journal
0 Comments