Google Discover filters stale articles before ranking: what the SDF paper shows

A research paper from Google describes a filter that removes articles from Discover before the ranking stage. Abondance analysed the paper, titled “Decomposing Staleness in Recommender Systems”, which six researchers will present at the CIKM conference in Rome from 7 to 11 November. It describes SDF, a system deployed in Discover for two years. Over that period, reader reports of stale content dropped by 54.9 percent. The paper describes Discover as it worked in the first quarter of 2026, and its examples are American.

Two ways an article goes stale

SDF treats staleness as two distinct problems.

The first is replacement. A newer article about the same event or entity contradicts, overtakes or outdates an earlier one. A first model compares articles in pairs to detect these cases.

The second is wear. The topic has a short life, and the article has already received most of its audience. A second model reads text, images and video and estimates the share of lifetime traffic the article has already received. At 90 percent, the article is withdrawn. In the paper’s example, a guide to the Perseid meteor shower crosses that threshold the day after the event. A gardening guide that draws steady visits over a month never reaches it.

One model is enough to remove an article. The filter runs before ranking, so strong engagement, popularity or a well-known publisher name cannot rescue an article once it has been flagged.

What publishers can adjust

The paper gives no advice to publishers. The six reflexes Abondance proposes are its own deductions from the mechanism, and should be read as such.

Publish the update. On an evolving story, the article that gives the outcome replaces the earlier ones. A title that states the result signals new information, which a generic recap title fails to do.

Keep a precise angle. A newer piece only replaces yours if it covers the same information. A broad overview of a market does not supersede a focused article on one specific point, such as fixed mortgage rates.

Keep time markers for news content. The model reads the content and has no access to your analytics. Words such as “tonight” or “live” in a title, or a photo of an event that is over, can push a durable guide into the short-life category. Evergreen pieces need undated titles and images.

Put less weight on in-house promotion. The model learns from Google Search clicks. The authors excluded page views, which they consider distorted by publishers promoting their own content. Direct traffic from homepages, notifications and social networks does not count toward an article’s estimated lifetime.

Avoid republishing. Reposting the same headline under a new URL restarts the clock, and in Abondance’s own data the topic stays visible about 50 percent longer. Abondance still advises against it: Discover demotes sites that do it at scale, particularly on expired domains bought back, and the visibility gained can expose the whole site.

Measure your own curve, which is the subject of the next section.

How to check your own traffic

Abondance notes that since July, large Discover articles keep roughly their best day and lose the following days. An end-of-life filter could produce that shape, but SDF was already running earlier in the year, so nothing ties this summer’s decline to it.

The test it suggests takes less than an hour. In the Discover tab of Search Console, compare the daily curves of your ten best articles from September and from February. If day one is comparable and the tail has disappeared, the pattern described applies to your site.

For acquisition teams, the practical consequence concerns forecasting. On news topics, Discover traffic is front-loaded and capped by an estimate of remaining interest. Evergreen content with clean, undated packaging keeps a better chance of a long tail.

For monetisation, this matters wherever revenue depends on page views per article, as with display advertising. A team can segment Discover revenue by content type over the last two quarters and check whether evergreen formats hold their value longer than news. That split shows where editorial hours return the most, and which formats deserve a refresh rather than a rewrite under a new URL.