Juan González Villa

Everyone talks about Google Discover like it's a black box. Post great content. Use big images. Hope for the best. That's the advice. That's been the advice for years. I decided to take a different approach. Instead of guessing, I went into Google's own SDK and read what it actually exposes during normal operation. Event constants. Telemetry counters. Configuration values. Things that are sitting right there if you know where to look. What came out was a picture of a 9-stage content pipeline, and it changed how I think about Discover entirely. Knowledge Graph entities? I need to look carefully! Here is the part that surprised me most: the collection-level filter runs before interest matching. Before the pCTR model. Before any ranking happens at all. If a publisher is blocked at that stage, Google never even evaluates whether their content matches what the user wants. It is binary. One article generates enough negative feedback, and the entire domain gets suppressed. There is no equivalent blanket boost. The penalty surface is wider than the reward surface. Some other things the telemetry confirms: og:title is not just a display label, it feeds directly into the predicted click-through rate model. The first 7 days carry the highest freshness weight. After 30 days, continuous staleness decay begins, tracked in hours. And dismissed content never comes back, tombstones are permanent records stored on-device. There are also 276 event constants, 150 concurrent A/B experiments running on a single session, and a system called NAIADES that most publishers have never heard of. The full breakdown, with every finding traced to a specific string or event, is on my blog. I also created a dashboard with all specifications with a FREE pCTR tool. Link is in the first thread!