Google has published research on the system it has been using in Discover for the past two years to predict and remove stale content before it reaches users. The system, called Supersession-Decay Filtering (SDF), relies on two AI-powered filters that work together to decide whether content remains useful. If either filter flags a page as stale, that content is blocked from ranking in Discover entirely, regardless of publication date or engagement history.
Search Engine Journal reports that the system has been live for two years, serving hundreds of millions of daily users, and that user complaints about stale content have declined by 54.9% since deployment. The technical detail matters because it changes how you should think about content lifespan, interlinking and the distinction between evergreen material and time-limited coverage.
How the two filters work, and why publication date does not decide freshness
Supersession-Decay Filtering separates two distinct ways content becomes stale. The first, supersession, happens when newer material covers the same topic better and makes the older page obsolete. The second, relevance decay, happens when content loses usefulness over time even if nothing has replaced it.
Traditional freshness systems used publication date cutoffs and engagement metrics. Google found both unreliable. A page can go stale days after publication if something newer supersedes it, while genuinely evergreen content might remain useful for years. Publication dates measure when something was published, not whether it is still relevant. Engagement metrics are reactive, interpreted long after users have already decided content is no longer useful.
The supersession filter evaluates the relationship between older and newer content to detect whether earlier information remains useful. The relevance decay filter predicts when content will lose its usefulness, even if nothing newer has replaced it. Together, they act as a filtering layer before content reaches the ranking system. If either filter identifies a page as stale, it is removed from the candidate set and does not reach users.
What this means for content that has a limited window
For most publishers, this shifts the question from "how often should we publish?" to "how long will this remain relevant, and what happens when it is not?" Breaking news has a short window. Product launches, regulatory changes and event coverage all have a defined shelf life. Once that window closes, the supersession filter will catch newer coverage and the decay filter will predict loss of usefulness.
The implication is that you need a plan for what happens after that window. The options are limited: update the content to reflect new information, redirect it to a newer page that supersedes it, or accept that it will be filtered out and stop expecting traffic from Discover. Letting old news sit unchanged is the worst choice, because the system is designed to remove it.
Evergreen content, by contrast, is what sustains a site when time-limited material decays. The research does not define how much of a content mix should be evergreen, but the logic is clear: if everything you publish has a short relevance window, you are constantly replacing traffic rather than compounding it. A higher proportion of genuinely evergreen material means a larger base of content that can survive both filters over time.
Where interlinking becomes a freshness problem
Many interlinking systems, whether plugin-based or AI-assisted, rely on semantic relevance to suggest links between related pages. Google's Supersession-Decay Filtering adds a dimension those systems do not account for: how long the linked content will remain relevant. Linking to a page that will be filtered out in a few weeks weakens the value of the link and, in aggregate, weakens the usefulness of the article doing the linking.
The practical answer is to prioritise interlinking to evergreen content where relevance will not decay quickly, and to treat links to time-limited material as temporary. That does not mean you should never link to breaking news or timely coverage, but it does mean you should expect those links to need replacing as the linked content becomes stale. Updating internal links after the fact is rare, which is why planning for it matters.
This also affects how you should think about cornerstone content. If a pillar page links heavily to time-limited articles that decay or get superseded, the pillar itself becomes harder to maintain. Structuring content so that evergreen hub pages link primarily to other evergreen material, and time-limited coverage sits in a separate layer, makes the job of keeping the site fresh over time more manageable.
The distinction between evergreen, limited-window and news content
The research paper, titled Decomposing Staleness in Recommender Systems: A Dual-Filter Framework for Supersession and Decay, does not prescribe a content mix, but it does make clear that publishers need answers to three questions: what percentage of content is genuinely evergreen, what percentage has a limited relevance window and needs a plan for updating or replacing, and what percentage is breaking news with a short lifespan.
Evergreen content is the material that will sustain the site with traffic for queries that remain useful and informative to users over time. Limited-window content covers topics that are relevant now but will be superseded or decay, and the plan for what happens when that window closes matters more than the decision to publish it in the first place. News content has the shortest lifespan and the clearest supersession risk, but it is also the most likely to drive spikes in engagement while it is still fresh.
The balance between those three types depends on the site. A news publisher will have a higher proportion of short-lifespan content and will need systems to update, archive or replace it efficiently. A service business or professional site can lean more heavily on evergreen material, with time-limited content used strategically rather than as the default.
What to track, and what to stop assuming
If Supersession-Decay Filtering has been live for two years, the system has already been acting on your content. The question is whether you have been tracking how long material remains relevant after publication, or treating freshness as a binary decision made once at the point of publishing.
Tracking relevance lifespan means recording when content stops driving traffic from Discover, when newer pages on the same topic get published, and when topics naturally lose search interest. That is not the same as tracking publication date or engagement rate. It is closer to content decay analysis, which most teams do not run because they do not have a reason to. This research gives you one.
The other assumption to drop is that high engagement or a recent publication date will protect content from being filtered. The supersession filter does not care when something was published if newer material has replaced it, and the decay filter does not care how popular something used to be if relevance has declined. Both are predictive, not reactive, and both act before content reaches the ranking system.
For sites that rely on Discover traffic, this is a forcing function. You either build a content operation that accounts for how long material will remain relevant, or you accept that a growing proportion of what you publish will be filtered out before it reaches users. The system is already live, and has been for two years. The catch-up is on your side, not Google's.