A research paper by six Google researchers, filed on arXiv on August 16, 2026, describes SDF (Supersession-Decay Filtering), a filter deployed in Google Discover. It removes content that newer coverage of the same topic has made stale, and content that a model predicts, from its text and images, has run out its useful life. Content that is removed never reaches the ranking stage. Over two years in production, user reports of stale content fell by 54.9%. The paper cites a durable explainer that the filter lets through, and a one-evening guide it drops.
Contents
Discover is a feature of Google Search that shows people content related to their interests.1 On August 16, 2026, six Google researchers filed a paper on arXiv describing how Discover removes stale content before ranking it.2 Search Engine Journal picked it up on October 21, seeing consequences for content and internal linking strategies.3 This article covers how the filter works and the limits its authors report, then what it means for a site that publishes news or dated articles.
01What SDF removes before ranking
The paper is titled “Decomposing Staleness in Recommender Systems: A Dual-Filter Framework for Supersession and Decay.” Its six authors are affiliated with Google, and it was accepted at the CIKM 2026 conference, held in Rome in November.2 It presents SDF as a system fully deployed in Discover.2
Two filters, one exit rule
The first filter handles supersession: content becomes stale when newer content moves the same topic forward. The second handles relevance decay: content gradually loses its value even when nothing replaces it.2 Each one relies on a learned model.2
| Filter | What it detects | How | Case cited in the paper |
|---|---|---|---|
| Relational (supersession) | content made stale by newer content on the same topic | a model that compares pairs of items | a “coming soon” post about the Mazda 3, dropped once its prices were revealed |
| Intrinsic (decay) | content nearing the end of its useful life | a model that predicts, from the content, the share of traffic already collected at several horizons | a meteor shower guide, removed after the event |
The two filters are combined by disjunction: it is enough for either one to flag an item for it to be removed, and an item that clears both stays eligible for recommendation.2 Search Engine Journal sums it up this way: if either filter identifies stale content, that content becomes ineligible for ranking in Discover.3
Upstream of ranking
SDF acts before ranking. It prunes stale candidates ahead of the most expensive step in computing terms, which the paper describes as still engagement-based and trend-aware.2 The signals are those of the item: neither filter depends on the user who is browsing, or on any context other than time.2
The decisions are computed off the request path and cached, so when the feed is served, applying a filter amounts to looking up a value that already exists.2 The supersession filter runs in two modes: an online mode, which labels news-related items in near real time, and a batch mode, which backfills the rest and extends a stale label to other items in the same cluster.2
The published results
Over two years in production, user-filed reports of stale content fell by 54.9% compared with the pre-deployment baseline: by 64.0% for supersession and by 34.4% for decay.2 Search Engine Journal repeats the 54.9% figure.3 A two-month experiment in which both filters were disabled for a group of users also measures a 0.16% drop in the rate at which users dismiss items and a 0.26% rise in positive engagement with the feed, with 95% confidence intervals of ±0.07 and ±0.18.2
The paper also quantifies the computing savings, monitored in the first quarter of 2026: 4.76% of serving CPU and TPU hours for the relational filter, and another 3.98% for the intrinsic filter.2
What publishers report these days
Search Engine Roundtable published an article by Barry Schwartz on October 21 about volatility in Google's rankings on October 20 and 21. He mentions an unconfirmed Google ranking update.4 Among the publisher comments he quotes, taken from his site and from WebmasterWorld, some concern Discover. One writes that Discover sends traffic for an hour or two and then dies, that stories get replaced by rivals, and thinks Discover's shelf life for content has changed: in that commenter's view, publishers must keep producing content to keep getting traffic.4 Barry Schwartz presents these comments as a sampling of industry chatter.4
02How Discover decides content is stale
The paper faults the usual methods for reflecting loss of relevance poorly. Ranking by recency demotes content instead of removing it, and users still have to scroll past it; a hand-tuned age cutoff is coarse, since a long one still serves stale items and a short one drops useful evergreen content.2 Search Engine Journal notes that a publication date only tells when the content was published.3
Supersession: one piece replaces another
The relational filter compares an existing item with a recent one on the same major topic, such as the same event or entity, and decides whether the new one makes the old one stale.2 Only updates that contradict the earlier item, supersede it or shift the narrative count, so the publication date alone cannot settle the question.2
The paper illustrates the case with news. An early bulletin on Queen Elizabeth II's health was superseded within hours by the announcement of her death, and a report that Shohei Ohtani was weighing offers from several teams became stale the moment his signing with the Dodgers was announced.2
To produce its labels at scale, Google has a large language model annotate more than a million pairs of items, then transfers that work to a T5-type model built for production.2
Decay: a lifespan predicted from the content
The second filter relies on the predicted traffic ratio, or PTR. For a given item and horizon, it is the share of the item's total traffic that has accumulated between publication and that horizon, the total being computed over a 30-day reference window.2 A news item collects most of its visits shortly after publication, while an evergreen item spreads them across its whole life.2
The model predicts this ratio from the item itself (text, image and video), without waiting for engagement to be observed.2 It is trained on visits from search click logs rather than page views, which the authors say are skewed by publisher promotion.2 The paper uses seven horizons, from 12 hours to 14 days.2
| Horizons | Life cycle targeted |
|---|---|
| 12 hours | breaking news, on an hourly scale |
| 1 to 3 days | daily reporting |
| 5 to 7 days | weekly content |
| 10 to 14 days | pieces whose interest decays more slowly |
A PTR at 7 days close to 1 indicates an item expected to collect most of its traffic within a week and little afterward. A low value indicates an item still drawing traffic at that horizon.2
What the filter lets through, and its limits
The paper shows production cases. A meteor shower guide meant for a single evening is dropped after the event, a guide to a multi-day food festival is kept during its dates, and a durable New York Times explainer on health coaching is not filtered.2
The authors also list the cases the filter misses. The supersession filter struggles with political commentary, where opinion and fact mix and no single item cleanly supersedes another. The PTR model misses a long tail of gradually decaying items, some of which become moderately stale without being caught. A substantial share of the remaining reports concerns content the user has already seen elsewhere, outside SDF's scope.2
03What follows for a site that publishes news
Google describes Discover traffic as less predictable than keyword-driven search visits, to be considered supplemental.1 It adds that sites may see that traffic change for reasons unrelated to the quality or publishing frequency of their content.1 From the paper, Search Engine Journal draws questions for an editorial site: the share of durable content, the plan for updating content that goes stale, and internal linking.3
Update the existing article rather than create a new page
The relational filter compares an item with the more recent ones on the same major topic.2 On a topic that keeps moving, each new article is a more recent item that can make the previous one stale. Updating the original article, with its title, introduction and figures, keeps a single current page for the topic.
Search Engine Journal asks whether the publisher has a plan for updating content that goes stale.3 A real revision of the page also shows up in the sitemap, through the URL's modification date.
Keep a share of durable content
The paper shows a durable item that the filter does not drop, and PTR is designed to tell content that collects its traffic in a few days from content whose visits are spread over time.2 Google's Discover page states for its part that older content may appear if it is helpful and relevant to individuals.1
Search Engine Journal raises the question of the share of evergreen content, which sustains the site with traffic for useful queries, the share with a limited window of relevance, and the share of news with a short shelf life.3 To set those shares, the page-by-page audit of an SEO checklist helps prioritize the texts to rewrite.
Link to the current version
Search Engine Journal notes that how long content stays relevant can guide internal linking decisions: many plugins and AI-based tools suggest links based on semantic relevance, and SDF adds freshness to that calculation.3
Links in older articles then point to the current version of the topic. Internal linking maintained over time means revisiting older articles with each new publication.