Google Discover Filters Stale Stories Before Ranking, Paper Says; User Staleness Reports Fell 54.9%

Ras Al Khaimah, United Arab Emirates, 23 October 2026.

Google Discover removes stale articles from its candidate pool before its ranking systems ever score them, according to a research paper by six Google engineers posted to arXiv on 16 August 2026. The paper describes SDF (Supersession-Decay Filtering), which it says is fully deployed in Discover, and reports that user-filed staleness complaints fell 54.9% over two years of production use.

The paper, accepted for the ACM CIKM 2026 conference in Rome on 7 to 11 November, drew wider attention among SEOs this week after Search Engine Journal reported on it on 21 October. It is the most detailed public account yet of how Discover decides that a story is out of date. Google has not announced SDF on its Search Central blog or in its Discover documentation; the description below comes from the paper itself.

What the paper describes

The authors split staleness into two separate problems and built one filter for each.

• Supersession: an article becomes stale because a newer one moves the story on. The paper's examples include a report that a baseball player was weighing offers, made stale by news that he had signed. A model compares pairs of related articles and decides whether the newer one makes the older one outdated.
• Relevance decay: an article loses value over its own lifespan even if nothing replaces it. A guide to "tonight's meteor shower" is the paper's example. A second model, called predicted traffic ratio (PTR), reads the article's text, images and video and forecasts what share of its lifetime visits it will have collected by a given age.

If either filter flags an article, it is removed from the candidate set and never reaches the ranker. Articles that pass both stay eligible. Ranking itself remains engagement based and trend aware, the paper says; SDF sits upstream of it.

Filter	Problem it targets	How it decides	Fall in user staleness reports
Relational staleness filter	Supersession by newer coverage	Pairwise model comparing an existing article with a newer, related one	64.0%
Intrinsic staleness filter (PTR)	Relevance decay over an article's lifespan	Multimodal model predicting the share of lifetime traffic already earned	34.4%
SDF combined	Both	Article removed if either filter fires	54.9%

Key findings

• Staleness reports fell 54.9% over two years against the pre-deployment baseline, with supersession reports down 64.0% and decay reports down 34.4%.
• Training data came from an LLM. The team used PaLM 2-L to label 1,100,133 article pairs, then distilled those labels into a T5 student model fine-tuned from an 11B checkpoint. The best configuration scored an F1 of 84.79 on the test set.
• Engagement held or improved. In a two-month holdback test, the authors report lower user dismissal rates, higher feed diversity and higher positive engagement with SDF switched on. Configurations were tuned so that engagement would not fall.
• It saves compute. Because stale items are dropped before dense scoring, the paper reports CPU and TPU savings in Q1 2026 from both filters.
• It has blind spots. The authors say the pairwise model struggles with political commentary, where opinion and fact mix, and PTR misses a long tail of articles that decay slowly and never cross its threshold.

Why age cutoffs were not enough

The paper argues that the usual industry tools handle staleness poorly. Fixed age cutoffs either keep stale items (if long) or drop useful evergreen content (if short). Engagement floors such as minimum click-through rates only learn an item is stale after users have been shown it repeatedly. Ranking by recency demotes old items rather than removing them, so users still scroll past them.

SDF's answer is to judge staleness from content and context rather than publication date alone. In the paper's production examples, PTR dropped a one-night meteor shower guide after the event, kept a multi-day food festival guide during the festival and left an evergreen health explainer alone.

Context

Discover is a large traffic source for publishers, and the paper describes it as serving hundreds of millions of daily and billions of monthly active users. Google's Discover documentation covers eligibility and content policies but does not describe a staleness filter. Google's Search Status Dashboard lists a February 2026 Discover update that ran from 5 to 27 February; the paper does not link SDF to that or any other named update.

Research papers do not always describe what runs in production today. In this case the authors state that SDF has been deployed for two years across successive model upgrades and locale expansion, and that they continue to upgrade it with newer LLMs.

What this means for SEO teams

For publishers and SaaS content teams that earn Discover traffic, the paper suggests that being made out of date by a newer article, including your own, can remove a story from Discover entirely rather than just pushing it down.

1. Treat follow-ups as replacements. When a story develops, a new article that supersedes the old one is likely to win. Update or clearly link the earlier piece so readers who land on it see the latest facts.
2. Know which content type you are publishing. Split your plan into short-lived news, content with a fixed relevance window (events, launches, deadlines) and evergreen material, and expect Discover exposure to end when the window closes.
3. Do not rely on republishing dates. The paper's point is that publication time is a weak proxy. Changing a date without changing the substance does not address supersession or decay as described.
4. Write for clear supersession signals. Headlines and openings that state the new development plainly ("X signs with Y") are the kind of text a pairwise model can compare.
5. Watch Discover in Search Console by content type. Compare how long news, event and evergreen pages keep earning Discover impressions to see where your own decay curves sit.

Conclusion

The SDF paper gives SEOs a rare, specific look at one Discover system: two learned filters, one for articles overtaken by newer coverage and one for articles past their useful life, applied before ranking. Google engineers report a 54.9% fall in user staleness complaints over two years, with engagement steady or better.

Google has not commented on SDF beyond the paper, and the paper does not say how it interacts with Discover's other ranking systems. For content teams, the clearest takeaway is that freshness in Discover is judged against the story and the article's expected lifespan, not just the timestamp.

Questions and answers

What is SDF in Google Discover?

SDF (Supersession-Decay Filtering) is a staleness filtering system that a Google research paper says is fully deployed in Discover. It removes stale articles from the candidate pool before the ranking stage.

What are the two filters?

A relational filter detects supersession, where a newer related article makes an older one outdated. An intrinsic filter, the predicted traffic ratio model, forecasts relevance decay from an article's content.

How much did SDF reduce stale content?

The paper reports that user-filed staleness reports fell 54.9% over a two-year deployment, with supersession reports down 64.0% and decay reports down 34.4%.

Has Google officially confirmed SDF?

The system is described in a paper by Google engineers posted to arXiv on 16 August 2026 and accepted at CIKM 2026. Google has not announced it on its Search Central blog or in its Discover documentation.

###