Google published new research on how to improve Google Discover with a more accurate system for identifying and removing stale content. The new system, Supersession-Decay Filtering (SDF), is already in use by Google and is affecting what content is shown to users.
Supersession-Decay Filtering (SDF) Is Already In Use
The word “supersession” is derived from the word “supersede” and identifying when newer content has superseded older content is one of the two things this system does. The word “decay” relates to the other thing the system does, which is to predict when content usefulness is lost through relevance decay.
What may be shocking to the publishing and SEO community is that this system has been in use for two years. That means publishers, site owners, and SEOs need to catch up. The details of this system have implications for content and interlinking strategies.
Key takeaways are:
- Uses two filters to proactively remove stale content.
- Once the stale content is removed the downstream ranking systems, which include user engagement signals and “trend-aware” systems, are able to rank web pages in Google discover.
The research paper explains:
“We address both mechanisms with SDF (Supersession-Decay Filtering), a unified staleness filtering system deployed in Google Discover, a personalized feed of open-web content in the Google Search app, serving hundreds of millions of daily and billions of monthly active users. While downstream ranking systems remain engagement-based and trend-aware, SDF acts as a proactive filtering layer. It prunes fundamentally stale content before it enters the computation-heavy ranking stage, providing a refined candidate set. SDF accomplishes this by decomposing the problem into two complementary components…”
The research paper goes on to describe two filters that are at the heart of the SDF.
Recommendation Engines And Freshness
Most people know what a search engine is. There are other kinds of engines and one of them is called a recommendation engine. Google Discover and YouTube are examples of recommendation engines.
Aside from matching content to user interests, a recommendation engine surfaces fresh content. Freshness is a key ranking factor for Google Discover, it’s what users prefer. The researchers discovered a way to proactively predict which content is no longer fresh and use that information to prune stale content before it reaches users.
Why Publication Date & User Engagement Metrics Fail
Traditional methods for identifying stale content rely on publication date cutoffs and user engagement metrics that indicate people stopped engaging with a topic.
The researchers explained that publication date does not reliably indicate whether content is still relevant or useful to users.
- Content can become outdated shortly after publication when it’s superseded by newer information.
- Content can also become stale because of Relevance Decay, which is when its usefulness declines over time, even when it’s not superseded by fresher content.
- Another problem with using cutoff dates is that relevant and useful evergreen content might be prevented from displaying.
- Publication dates only measure when something was published and not whether it’s relevant or if it’s been superseded.
User Engagement Metrics
User engagement metrics fall short because engagement-based systems are reactive. The signals are interpreted long after users have already decided that content is stale.
Supersession-Decay Filtering (SDF)
The word “supersession” is derived from the word “supersede” and superseding stale content with fresher content is what this system is all about. The researchers developed Supersession-Decay Filtering (SDF) as way to proactively detect when content is obsolete because of newer information and to predict when the content naturally loses relevance.
- Supersession
The first filter identifies content that has become outdated because newer information has superseded it (supersession). The AI evaluates the relationship between older and newer content to detect if the earlier information remains useful. - Relevance Decay
The second AI-powered filter evaluates content to predict when it will lose its usefulness, even if nothing newer has replaced it. Another AI-powered filter predicts when content will lose relevance.
Together, they remove stale recommendations before they reach the ranking system, improving content freshness while also reducing computing costs.
SDF Blocks Content From Ranking
The two filters (Supersession and Relevance Decay) work together as a single system. If either filter identifies stale content then that content is made ineligible for ranking in Google Discover. Content that can make it through both filters remain eligible.
The Results Of Testing
The researchers found that SDF reduced stale content recommendations compared to traditional methods. User engagement showed modest gains, improving user engagement and content diversity. Over two years of deployment, user complaints about stale content declined by 54.9%. The system also reduced computing costs by removing outdated content before ranking.
Takeaways
The research paper is titled, Decomposing Staleness in Recommender Systems: A Dual-Filter Framework for Supersession and Decay. I believe that Google’s SDF makes it imperative for publishers have answers for:
- What percentage of content should be evergreen?
This is the content that will sustain the site with traffic for queries that are useful and informative to users. - What percentage has a limited window of relevance is there a plan for updating content when it goes stale?
- What percentage of content is news and has a short window of relevance?
Ideally it’s important to cover all breaking news that is relevant to an audience.
How long will content remain relevant is important because it can be used for decisions about interlinking articles. Many plugins and AI-based systems for interlinking content is based on semantic relevance. Google’s new SDF system adds the dimension of freshness to that calculation.
Featured Image by Shutterstock/JHVEPhoto
