Google’s Discover job posting confirms what the feed already shows

Google rarely explains how Discover really works. But sometimes it reveals more than intended—in a job posting.

Google recently listed a role for a Staff Software Engineer, Discover Ranking in Mountain View. One line in the minimum qualifications stands out:

• “5 years of experience building and deploying recommendation systems models (retrieval, prediction, ranking, embedding) in production”

The “Staff Software Engineer, Discover Ranking” posting on Google Careers, with retrieval, prediction, ranking, and embedding highlighted in the minimum qualifications. The four words sit together in the minimum qualifications of Google’s Discover Ranking posting (Google Careers, Mountain View). Highlight added by us.

Four words: retrieval, prediction, ranking, embedding.

A job posting describes a skill set. Nothing more. Its value is in the overlap.

We’ve monitored real Discover feeds for two years — 42 million cards and counting. Three of those four words match layers we’ve already traced from the outside. Here’s what that looks like, brick by brick.

Retrieval: the candidate layer has visible seams

Before anything gets ranked, the system has to decide which articles and videos are even eligible for your feed. We’ve mapped about 20 pipelines that feed Discover (full study), and several include “retrieval” in their internal names:

• A candidate sampling pipeline dominated by diversified editorial content.
• A post-retrieval evaluation pipeline made up almost entirely of YouTube and X content.
• Cluster-profile retrieval variants.
• A trend-embedding retrieval channel.
• Item-item collaborative filtering.

One number is worth remembering: a generative retrieval channel appeared as early as September 2025 in roughly 0.03% of the French Discover feed. That’s consistent with Google testing LLM-driven candidate selection on a tiny slice before any broader rollout.

The pipeline name is often the only evidence we have, so its exact role remains a matter of interpretation. But the layer itself is unmistakably there.

Prediction: the scores behave like the job posting says they should

Every Discover card we observe carries a set of prediction scores — probabilities between 0 and 1. We’ve isolated about nine of them, and they collapse into two nearly independent dimensions:

• An attention axis (will the user stop on this card? — mostly a property of the article).
• An engagement axis (will this specific user click and read deeply? — a property of the user-content pair).

The correlation between the two is close to zero. A headline can grab attention without holding it—the clickbait pattern in one sentence.

These scores track real behavior. In our panel measurements, observed interaction roughly doubles from the bottom to the top of the deep-engagement score scale, while steadily declining as the “scroll-past” score rises.

In other words, Google’s predictions closely match the behavior we later observe.

The biggest effects we measure come from personalization.

On our test accounts, we compared two sports publishers in the same vertical with nearly identical topic potential. One received deep-engagement predictions about 2x higher and final amplification roughly 8x stronger.

The striking detail is that the dominant publisher is the one fewer of our panel accounts follow through Google’s Follow feature. That suggests amplification is driven primarily by the affinity the model has already learned between readers and the source — beyond the topic itself and beyond the explicit follow.

A mirror test on a single tech publisher completes the picture. Accounts that follow it receive deep-engagement predictions nearly 2x higher than those of accounts that don’t, indicating the Follow button serves as one signal among several that shape that affinity.

A U.S. replication (ESPN vs. the NFL’s official site, NFL.com) shows the same pattern with a smaller gap: 1.28x amplification. (These are small samples; the full methodology and charts are in the complete study.)

Two sports publishers at nearly identical topic potential: the one the model predicts more engaging gets deep-engagement scores about 2x higher and roughly 8x more amplification.

On our test accounts, two French sports publishers (L’Équipe and FootMercato) at nearly equal topic potential. The one the model predicts more engaging gets deep-engagement scores about 2x higher and final amplification on the order of 8x, even though our panel accounts follow it less.

Small sample, an illustration rather than a general proof.

Embedding: several named families per user

To retrieve candidates and generate predictions, the system needs compact representations of users and content.

What we observe suggests Google maintains several named embedding families per user, each operating over a different time window:

• Discover interests (including a short-term variant).
• A trends-oriented family.
• A real-time family.
• A shopping-related family that also appears behind AI summary cards in finance and tech news.

These vectors appear to feed a classic two-tower retrieval system: one tower for the user, one for the content, with recommendations driven by proximity in a shared vector space. The names are observed; the roles are our interpretation.

What this means for publishers

The job posting validates the vocabulary. The feed shows the machinery at work. Three practical takeaways:

• Reader affinity is a direct distribution lever. At equal topic potential, the affinity the model has learned between readers and a source increased amplification by roughly 8x in our tests. It outweighed the explicit Follow signal, which appears to be just one input among several.
• Attention and engagement are separate diagnostics. They vary almost independently. Improving one doesn’t necessarily improve the other.
• Engagement is only half the story. The posting itself says “more engaging, and useful.” Usefulness appears to depend on signals beyond the prediction scores we’ve identified.

Google’s job posting names the four building blocks of its recommendation system. Three already leave measurable traces in real Discover feeds. The fourth — ranking — reveals itself every morning in what your readers see, or never see.