A single qualification line in Google’s recent “Staff Software Engineer, Discover Ranking” job ad did what the company rarely does: it named the platform’s core technical building blocks—”retrieval, prediction, ranking, embedding.” For publishers and marketers, that short list is more than vocabulary. Independent monitoring of roughly 42 million Discover cards shows three of those layers leave measurable traces in the feed, and each layer maps to distinct distribution levers publishers can influence.

Why that lone line matters

Google usually keeps Discover’s internals opaque. A public job posting that explicitly groups those four terms acts like a partial architecture note: it validates categories observers had already inferred from external signals. That matters because different problems—being considered as a candidate, grabbing attention, or holding a reader—are handled by different subsystems. Treating distribution as a single metric risks misallocating resources.

What the public feed shows, layer by layer

Retrieval. Before any ranking happens, the system selects eligible candidates. Monitoring uncovered around 20 distinct pipelines whose names and behaviors match retrieval stages: editorial sampling, post-retrieval evaluation channels dominated by YouTube and X, cluster-profile variants, trend-embedding channels and item-item collaborative filtering. In one test slice from September 2025, a generative retrieval channel appeared in roughly 0.03% of the French feed—consistent with a small-scale LLM-driven candidate test.

Prediction. Each Discover card carries multiple prediction scores—about nine can be isolated—which collapse into two largely independent dimensions. One axis predicts attention (will a user stop on the card?) and the other predicts deep engagement (will this specific user click and read deeply?). In our panel measurements, interaction roughly doubles from the bottom to the top of the deep-engagement scale and declines as a scroll-past score rises. Personalization amplifies these effects: in paired tests with two sports publishers at similar topical potential, the publisher the model predicted as more engaging received roughly 2× higher deep-engagement predictions and about 8× the final amplification.

Embedding. The feed behavior implies multiple named embedding families per user operating over different time windows: Discover interests (including short-term variants), a trends family, a real-time family and a shopping-related family that also surfaces behind AI summary cards in finance and tech. These vectors appear to power a two-tower retrieval setup—user vectors compared with content vectors in a shared space—to surface candidates.

Ranking. The posting’s inclusion of “ranking” confirms it’s treated as a discrete component alongside retrieval, prediction and embedding. Ranking is visible only indirectly—how cards are ordered and surfaced—but the job ad reinforces that ranking is an explicit stage, not an emergent artifact.

Practical implications for publishers and marketers

  • Prioritize affinity over chasing follows. The model’s learned affinity between readers and a source can drive amplification far more than the explicit Follow signal. In tests, the source favored by the model got substantially more visibility even when fewer accounts had tapped Follow. That suggests investments in consistent, recognizable signals of source affinity—habit-forming formats, repeat topical authority and persistent brand cues—will likely pay bigger distribution dividends than focusing solely on follow counts.
  • Separate attention and engagement metrics. Attention (stops, impressions) and deep engagement (click-through, time on page) are distinct diagnostics. Headline or thumbnail changes that boost attention won’t necessarily move deep-engagement predictions. Instrument both early and deep metrics, and run experiments that isolate which creative or UX changes shift each axis.
  • Signal usefulness, not just clicks. The job posting pairs “engaging” with “useful.” That implies Google’s distribution criteria include signals beyond raw engagement—recency, credibility, authoritativeness and vertical-specific indicators (for example, shopping signals). Publishers should surface signals of utility in metadata, structured data and content templates where appropriate.
  • Optimize by layer, not by a single KPI. Discover’s stack is multi-layered. Work to increase retrieval eligibility (how content gets surfaced as a candidate), cultivate affinity (signals that link users and source), and tune elements that influence attention versus deep engagement separately. Different teams will own different layers: editorial for candidate relevance, product/UX for friction and engagement, and data science for affinity signals and measurement.

These recommendations are grounded in observational data and inferred pipeline roles rather than Google’s internal diagrams. The job posting supplies a useful taxonomy that aligns with measurable behavior, but some pipeline names and exact functions remain interpretive.

What to watch and do next

Publishers should track three developments closely: whether generative retrieval channels expand beyond tiny test slices; shifts in the weight Google assigns to explicit follows versus inferred affinity; and the emergence of new embedding families tied to shopping or real-time trends. Practically, start by splitting optimization work by layer—retrieval eligibility, attention, engagement and usefulness signals—and instrument tests that measure movement on each prediction axis rather than relying on an aggregate distribution metric.

Google’s job posting doesn’t disclose implementation details. It does, however, give publishers a clearer map of the system they’re trying to influence. Use that map to target eligibility, build affinity and optimize the separate signals that make content both attention-grabbing and genuinely useful.