DeepTrends and DeepTrendsFable: Discover's trends layer (5% of volume, 27% pass rate)

Discover's Layer 3 handles emerging trends. Two pipelines compose it: DeepTrendsFable detects topics that are rising, DeepTrends persists those that hold. It's a brutal sequential detector — only 27% of topics detected at the first stage pass to the second. Here's what the public data actually says.

Where what we know comes from

Two main sources converge on this layer:

1. 1492.vision (Sylvain Deauré & Damien Andell) — analysis of 42 million Discover cards between December 2025 and February 2026. This is the source that identified and named the DeepTrends and DeepTrendsFable pipelines, measured their volume share (~5% combined) and observed the 27% pass rate between the two stages. It's also this source that documents DeepTrends as contracting in the French market.
2. Metehan Yesilyurt — metehan.ai — decompilation of the Android Google App SDK (87,498 Java classes). Confirms via code-source angle the existence of these pipelines observed behaviorally by 1492.vision.

SearchEngineLand relayed part of the 1492.vision analysis in February 2026 and specifies an additional parameter: 21h delay measured between detection by DeepTrendsFable and persistence by DeepTrends.

The May 2024 Content Warehouse leak, by contrast, does not mention these pipelines by name. Mike King (iPullRank) doesn't cite them in his analysis. Everything that follows therefore rests on empirical feed observation and SDK decompilation, not on internal Google documentation.

DeepTrendsFable — the detection stage

DeepTrendsFable is the emerging-topic capture stage. Its precise internal function isn't documented by Google, but observation shows:

• X.com (formerly Twitter) is explicitly identified as a trend signal source — including in the English feed. Notable because X is often considered a North-America-dominant signal.
• 1492.vision's analysis of the creatorcontent pipeline confirms that x.com feeds 75% of the creator flow in FR (vs YouTube 72% in EN). It's a detection infrastructure clearly oriented toward social text in France.
• Other signal sources are plausible (search growth, editorial publication velocity, engagement signals) but aren't confirmed by a primary source. Treat them as hypotheses.

DeepTrends — the persistence stage (contracting)

DeepTrends is the stage that persists detected trends from DeepTrendsFable — keeping them visible in the feed after the detection phase. Again, the precise internal function isn't documented.

Important data point from 1492.vision: DeepTrends is contracting in the French market over the observation period. Its volume share is decreasing. This is consistent with a broader evolution 1492.vision observes: Discover progressively shifts from query-based pipelines (legacy) toward pipelines based on embeddings and real-time personalization (Nephesh, Picasso, VanGogh). The pure trends layer loses relative importance in favor of continuous personalization.

The figures to remember

Three measured parameters for this sequential layer:

1. Combined volume: ~5% of feed (1492.vision). Modest in volume but strategic on timing.
2. Pass rate: 27% (1492.vision). Of topics detected by DeepTrendsFable, ~1 in 4 passes the persistence step by DeepTrends. The other 3 are dropped. It's the key filter deciding who stays visible.
3. Delay: 21 hours (SearchEngineLand). Time measured between detection and persistence. Important: this is not an editorial recommendation (« publish 21h before peak »). It's an internal pipeline parameter. We can infer that editorial reactivity matters, but no source specifies the exact pre-peak window during which publishing is optimal.

What this means for a publisher

1. Publish early in a trend's cycle. Persistence follows detection — being present when DeepTrendsFable picks up a topic increases the odds your articles are the candidates DeepTrends persists.
2. Monitor X.com as a detection signal. Confirmed by 1492.vision and SearchEngineLand. If a topic takes off on X in your niche, it's a trend indicator likely to be picked up by DeepTrendsFable.
3. Don't bet everything on this layer. Only 5% of volume, and DeepTrends contracting. The trends layer complements other pipelines (content, moonstone, aura) — it doesn't replace them. An article passing DeepTrendsFable but failing elsewhere won't drive big volume.
4. Measure on your feed. Search Console > Performance > Discover remains the only ground truth for spotting which trends actually work for you. Our Topic Radar tool helps spot early-rising topics in your category.

What we removed (transparency)

Previous versions of this article contained unsourced claims: « publishing 12h before peak = 5x traffic », « 3 days late = crumbs », or describing DeepTrendsFable as a « narrative builder that picks publishers ». Those claims sounded sourced but weren't. We removed them in favor of what's actually attested by public sources.

Sources

• 1492.vision — Sylvain Deauré & Damien Andell (primary source: 42M Discover cards analyzed Dec 2025 - Feb 2026, 27% pass rate, DeepTrends contracting in FR)
• Metehan Yesilyurt — metehan.ai (Google App SDK decompilation, 87,498 Java classes, independent validation of the architecture)
• SearchEngineLand — Inside Google Discover: 20 pipelines, 42 million cards (relay with the 21h delay)

Frequently asked questions