TL;DR
- The paper, arXiv 2608.15780, was posted on 16 August 2026 by six Google LLC authors and accepted to CIKM 2026.
- Supersession reports fell 64.0% and decay reports fell 34.4% over the two-year Discover deployment.
- One filter compares pairs of items to spot overtaken coverage; the other predicts each item's share of lifetime traffic.
- The paper makes no claim about Google Search and does not say whether editing an article affects how SDF treats it.
A research paper by six authors at Google LLC says Google Discover runs a filtering system called Supersession-Decay Filtering (SDF) that removes stale articles from the candidate pool before Discover's ranking systems ever see them. According to the paper, SDF has been in production for about two years, and over that period user-filed staleness reports in Discover fell 54.9% against the pre-deployment baseline. This is not a new update. It describes a system that has been running for some time, and Google has made no separate announcement about it.
What the Google Discover staleness paper says
The paper is titled "Decomposing Staleness in Recommender Systems: A Dual-Filter Framework for Supersession and Decay". It was posted to arXiv as paper 2608.15780 on 16 August 2026 and accepted to the CIKM 2026 Applied Research Track, the ACM conference on information and knowledge management held in Rome from 7 to 11 November 2026. All six authors (Di Bai, Feng Han, Zhenwei Tang, Jintao Liu, Luoshu Wang and Jialu Liu) list Google LLC as their affiliation.
The abstract describes SDF as "a staleness filtering system fully deployed in Google Discover", which it calls a personalized feed with "hundreds of millions of daily and billions of monthly active users". The full text calls Discover "a personalized feed of open-web content in the Google Search app". Search Engine Journal's Roger Montti brought the paper to wider attention in an article on 21 October 2026.
Two ways an article goes stale
The paper splits staleness into two mechanisms, and SDF has one filter for each.
Supersession
Supersession is when newer coverage makes an older item out of date. The paper's examples are an early report that Queen Elizabeth II was under medical supervision, overtaken within hours by news of her death, and a report that Shohei Ohtani was weighing offers, overtaken once his Dodgers signing was announced. The paper stresses that this signal is relational: an old item becomes stale only when an update contradicts it, supersedes it or moves the story on, so detection means comparing pairs of items rather than only checking publication times.
Relevance decay
Relevance decay is when an item simply runs out of usefulness, with no follow-up involved. The paper's examples are a guide to "tonight's meteor shower", which matters for one evening, and a local feel-good clip that fades on its own. Here the signal belongs to the item itself and its expected lifespan.
How the two filters work
The relational staleness filter is a pairwise classifier. Given an existing item and a newly arriving one, it predicts whether the new item makes the old one stale. Because labelling pairs by hand is too slow at this scale, the team used an LLM, PaLM 2-L, to label synthetic training data, producing labels for 1,100,133 examples, and then distilled that into a smaller T5-based student model. The student also writes a short rationale before its answer. In the paper's offline tests, the deployed configuration reached an F1 score of 84.79.
The intrinsic staleness filter is called the predicted traffic ratio (PTR) model. It reads an item's text, images and video and predicts what share of its lifetime traffic it will have collected by a given point after publication. A breaking news item collects most of its visits quickly. An evergreen piece collects them slowly over its whole life. When the predicted share for an item's current age passes a threshold, the item is treated as having had its useful life and is removed. The paper says PTR's training labels come from search-click logs rather than raw page-view logs, because page views are skewed by publisher promotion, and that labels are calculated against a 30-day total.
The two outputs are combined with a simple OR: if either filter flags an item, it leaves the candidate set. The downstream ranker, which the paper says stays engagement-based and trend-aware, only ranks what remains, so "removed items are never served". Both filters use item-level signals that do not depend on the user, and they are computed ahead of time, so serving needs only a lookup.
The numbers in the paper
The figures below are the ones the paper reports for SDF's deployment in Google Discover.
| Measure | Result reported in the paper |
|---|---|
| Total user-filed staleness reports, two-year deployment | Down 54.9% against the pre-deployment baseline |
| Supersession reports | Down 64.0% |
| Decay reports | Down 34.4% |
| LLM-labelled training pairs for the relational filter | 1,100,133 examples, labelled by PaLM 2-L |
| Relational filter offline F1 (deployed configuration) | 84.79 |
The paper also describes a two-month holdback test, in which SDF was switched off for a group of users. Against that group, SDF reduced the rate at which users dismissed items and lifted feed diversity and positive engagement. Filter settings were tuned to be "engagement-neutral", which the authors say rules out simply cutting more content to make the staleness numbers look better.
Where SDF gets it wrong
The authors list their own failure cases, which are as useful to publishers as the wins. The relational filter struggles with political commentary, where opinion and fact are mixed and no single piece clearly replaces another. PTR misses a long tail of items that decay gradually, because their predicted traffic share never reaches the threshold. And a large share of remaining staleness complaints come from content the user had already seen on Discover or elsewhere, which SDF does not try to handle.
The wins are concrete. The relational filter marked "Mazda 3 IPM launching soon?" stale once "Mazda 3 prices revealed" arrived. PTR dropped a Geminid meteor-shower guide after the event, kept a multi-day guide to a Phuket food festival while it was still relevant, and left an evergreen health-coach explainer alone.
What this is not
A paper about Discover invites over-reading, so the limits are worth stating clearly:
- It is not a new algorithm update. The paper describes a two-year production deployment, with model upgrades and expansion to more locales during that time.
- Google has not issued a statement about SDF beyond the paper itself.
- The paper is about Discover. It does not say SDF, or anything like it, is used in Google Search rankings, and nothing here should be read as a Search ranking factor.
- The paper does not publish its thresholds, and the PTR horizons are described only in broad terms, from hours for breaking news up to two weeks for slower content.
- It does not say whether updating an existing article, or changing its modified date, affects how SDF treats it. The paper takes an item's arrival time from when it is annotated in the indexing pipeline, which it says closely approximates publication time.
What publishers can take from it
Analysis: the points below are inferences from the mechanism the paper describes, not findings the paper reports.
Follow-up coverage replaces earlier coverage
If a newer article on the same event can make an older one ineligible for Discover, a publisher's own follow-up will likely retire its earlier piece too. That is by design, and it means a story that develops over several days will not keep earning Discover traffic from its first version. Each meaningful development needs its own clear, complete piece if it is meant to compete.
Update in place or publish a new URL
The paper does not answer this, so it comes down to editorial judgement. Live-updated pages keep links and history in one place, but the paper ties an item's age to its indexing time and says nothing about resets for edits. Separate URLs for real developments, such as a result after a preview, map more directly onto the supersession model the paper describes. Publishers who rely on Discover can test both on their own sites and compare.
Plan shelf life by content type
PTR treats time-bound pieces, like event guides, as having a short life, and evergreen explainers as having a long one. Expect a short Discover window for dated content, and do not judge evergreen pages by their first-day Discover spike. Discover is one channel among several, which is why SEO for search visibility and Discover performance should be read separately in reporting.
Planning which pieces are time-bound and which are evergreen is part of content marketing strategy, and the paper gives a useful frame for that split.
What this means for Thai marketers
For Thai publishers and brands that get traffic from Discover, the useful takeaway is the mechanism, not a new rule. The paper says its PTR training data covers multiple languages and formats, and it mentions expansion to more locales over the deployment, but it does not list which languages or countries. One of its own examples, the Phuket food-festival guide, shows the system handling a multi-day local event. Thai news and lifestyle sites that publish many updates on one story should expect earlier versions to drop out of Discover once newer ones arrive, and should plan reporting with that in mind.
FAQ
Is SDF a new Google Discover update?
No, SDF is not a new update. The paper describes a system deployed in Google Discover over a two-year period, and Google has made no separate announcement about it.
Does SDF affect Google Search rankings?
The paper does not say so. It describes SDF only in Google Discover and makes no claim about Google Search.
How much did stale content complaints fall?
User-filed staleness reports in Discover fell 54.9% against the pre-deployment baseline over two years, according to the paper. Supersession reports fell 64.0% and decay reports fell 34.4%.
Will updating an old article bring it back into Discover?
The paper does not say. It takes an item's arrival time from its indexing-pipeline annotation time and does not discuss edits or modified dates.
Does this apply to Thai-language content?
The paper does not list languages or countries. It says the PTR training data spans multiple languages and that deployment expanded to more locales, and one example is a guide to a food festival in Phuket.
The bottom line
Researchers at Google have now described, in public, how Discover removes stale articles before ranking: one filter for coverage that has been overtaken, one for content that has run its course. On the paper's mechanism, that would explain why Discover traffic to a news piece can stop when a follow-up appears. It does not change anything today, and it says nothing about Search. If you want help reading Discover and Search performance side by side, the Relevant Audience team is glad to take a look.
