Ranking Algorithm

Overview

The pipeline runs in two stages. First, we cluster: items that report the same underlying story are grouped into a single cluster, no matter how many outlets covered it. Second, we rank: each cluster gets a composite score built from a set of structured signals about the story and its coverage. The top-signal clusters are what you see on the home Brief and on each topic and entity page.

This page describes the ranking side. For how we write cluster summaries, what sources we include, and what we do not do editorially, see Editorial Standards. This page focuses on the algorithm itself — what signals feed it, what it produces, and what it will not do.

Exact ranking weights and formulas are not published. The signals below describe the shape of the algorithm and the direction each signal moves a story, which is what most readers need to understand a rank and what publisher-quality systems look for as a methodology signal.

Signals

The composite score is built from six signal categories.

Source corroboration

How many independent publishers reported the same story. More independent sources raises confidence — several outlets converging on the same event is a stronger evidentiary signal than one outlet asserting it alone. Single-source stories are not excluded (a frontier lab publishing a paper alone is legitimate news) but they need to clear a higher bar on other signals to reach the top of the Brief.

Publisher authority tier

Sources are grouped into tiers by editorial policy. Frontier labs (OpenAI, Anthropic, DeepMind, Meta AI, Mistral, and similar), established outlets (major tech press with real reporting operations), and developer communities (arXiv, GitHub trending, and equivalents) each carry different weight. A single frontier-lab announcement can rank without wide coverage; a single small-blog post needs corroboration.

Headline signal strength

An LLM-assisted judgment of whether a headline is informative or promotional. Headlines that report a specific event score higher than headlines that recycle vendor talking points. This reduces our exposure to press-release SEO games, where every launch gets breathless coverage designed to fill the top of aggregator feeds.

Citation velocity

How quickly a story is picked up across the source pool. Rapid multi-source coverage over a short window is a strong signal of a real event. We weight citation strength alongside speed, so herd-following on rumor cycles does not rank as high as coverage of confirmed news.

Topic classification

Every cluster is classified into one of nine topics: paper, model_release, safety, policy, product, infra, funding, opinion, and other. Ranking is normalized within topic where it helps — a policy story competes against other policy stories rather than against a model release, so a slow news day for one topic does not crowd out its top story.

Story growth

How much a cluster's source set has expanded since the last score. Fresh corroboration in the current news cycle reweights ranking, so a story that keeps attracting new coverage is treated differently from a story that hit its peak yesterday.

Bucket classification

Every cluster is assigned a newsworthiness bucket — a coarse tier label like significant, research, tool, meme, and a few others. Buckets are derived from the composite score together with the story's topic and shape. The raw score sits underneath, but the bucket label is what most readers use to skim.

You will see bucket labels on cluster pages as small tier chips near the title. significant marks the day's most consequential stories. research marks papers and technical results. tool marks a new artifact — a library, a checkpoint, a demo. meme marks stories that are being discussed heavily but are not, in our editorial judgment, materially important.

Buckets are recomputed on every re-score, roughly hourly. A story can move up or down a tier as new sources land or as the composite score updates.

What the algorithm does not do

There are adjustments we deliberately do not make. This section exists so both readers and third-party quality systems can verify it.

• No paid placement. No source can pay to raise their ranking, appear more often, or unlock a tier chip. There is no such product and we will not build one.
• No competitive-source deprioritization. We do not downgrade a source because they compete with us or because they cover us critically.
• No hidden ranking pins. The only manual overrides are editor-pinned featured items, and those are visibly labeled as pinned wherever they appear.
• No personalization in the public feed. Every unsigned-in visitor to a given page at a given moment sees the same ranking. Signed-in paid users can opt into personalized ordering, but personalization only reorders the same set of clusters — it does not add sources, remove sources, or reweight them.
• No LLM-fabricated clusters. Every cluster on the site references real articles at real URLs published by real humans at real outlets. Nothing is synthesized to fill a slot.

Refresh cadence

The pipeline runs every 15 minutes. Each cycle it polls sources for new items, clusters new items into existing or newly created clusters, re-scores any cluster whose source set changed, updates bucket labels, and refreshes the Brief.

Cluster summaries and the prose on hub pages are generated by an LLM (currently Gemini 2.5 Flash). Ranking itself is not LLM-driven — it runs on structured signals. See Editorial Standards § AI use for the full AI-use disclosure.

The 15-minute cadence covers both the ranking recomputation and the front-end cache. Editor-pinned featured content is the only manual override on ranking, and it is labeled every time it appears.

Feedback

Disagreements about ranking are welcome. They are not, however, treated as corrections — the Corrections Policy covers factual errors in our editorial layer, not judgment calls about which stories rose or fell.

If a specific cluster feels ranked wrong, email [email protected] with the cluster URL, why the rank feels off, and any counter-evidence (widely-reported story we ranked low, single-source story we ranked high, misclassified topic). Concrete feedback of this shape is how we tune the signal weights over time.