Google posted a listing for a Staff Software Engineer, Discover Ranking, in Mountain View, and the minimum qualifications spell out four components most publishers have never seen named together: retrieval, prediction, ranking, and embedding. That single line is the closest thing to an org chart Google has ever published for how Discover decides what lands in a reader’s feed.
The listing matters because it corroborates independent measurement rather than replacing it. 1492.vision, a research firm that has tracked roughly 42 million Discover cards over two years, mapped around 20 internal pipelines feeding the surface, several carrying “retrieval” in their names, including a generative retrieval channel that showed up in about 0.03 percent of the French feed as early as September 2025.
On the prediction layer, the firm isolated nine separate scores that collapse into two nearly independent dimensions: whether a card grabs a reader’s attention, and whether that specific reader engages with it deeply once clicked. The near-zero correlation between the two explains a familiar pattern: a headline can pull a stop without earning a read.
Personalization produced the largest swings in the firm’s testing. Comparing two sports publishers with similar topical strength, the outlet with higher predicted engagement drew roughly 2x higher engagement scores and close to 8x more amplification, despite being followed by fewer of the panel’s test accounts. A parallel U.S. test pairing ESPN against NFL.com found a smaller but consistent gap, at 1.28x. That pattern suggests Google’s model has learned publisher-level affinity that operates independently of, and can outweigh, the explicit Follow signal.
The embedding layer, the study found, runs on multiple named vector families per user rather than one, including separate representations for short-term interests, trending topics, real-time behavior, and shopping intent, the last of which also surfaces behind AI summary cards in finance and tech coverage. Those vectors point toward a two-tower retrieval design, matching users and content by proximity in a shared vector space, a common architecture in large-scale recommendation systems but one Google has not previously confirmed for Discover.
None of this comes from Google on the record. The job posting confirms vocabulary, not mechanics, and the firm’s own findings are built on panel accounts and sampled feeds rather than Google’s internal logs, so the exact weighting between affinity, topic, and freshness remains an inference, not a disclosed formula.
For publishers, the practical shift is where to spend effort. If model-learned affinity between a reader and a source can produce an 8x swing in amplification independent of topic fit, then building a base of consistently engaged readers matters more than optimizing any single Discover-eligible headline. Teams chasing Discover traffic should track deep-engagement proxies, like return visits and read-through, alongside click volume, since the two dimensions the model scores separately are also the two a publisher needs to move separately.
Search Engine Land published this analysis on August 26, 2026, citing a Google Careers listing and two years of independent Discover feed monitoring.
