Google patent describes fan-out queries built from embeddings, not text

Roger Montti at Search Engine Journal surfaced a Google announcement describing a new framework for fan-out queries, which SEO practitioner Chris Long wrote up on LinkedIn on 26 September 2026. The framework is named R4T-Diffusion.

The system would train a model to produce fan-outs from embeddings rather than have an LLM write out textual search queries. Long quotes the filing's phrase about deriving "content embeddings directly from query embeddings".

Three things stand out in the filing. The model is trained against "groundedness, diversity, and alignment", with diversity aimed at stopping fan-outs that look alike. The reasoning can move to an offline system, which Google frames as cheaper and faster. That cost profile would let fan-out sets grow larger than they are now.

Long's own read, based on work at Nectivv, is that current Gemini fan-outs are basic, and this looks like Google's route to more diverse sets at lower cost.

What is and is not established

This is a patent-style filing, not a shipped change. We think the useful move for practitioners is to log the fan-outs you can currently observe against your own content, so that if diversity widens later you have a baseline to compare against. Verify on your own pages before drawing conclusions from any one example.