Google Discover’s New Feed Feature Highlights Niche Sites

Google has introduced a significant shift in how users can personalize their Google Discover feed, moving from implicit interest tracking to a system where you can simply type what you want to see. The feature, initially called “Tailor Your Feed” before rebranding to “Add topics to your feed” in spring 2026, allows users to shape their content stream using natural language prompts. This development marks the first time users can explicitly command the feed’s behavior rather than relying on Google’s inferred signals from clicks and dwell time. Our analysis, based on field tracking of the Google app since December 2025 and detailed pipeline observation, reveals ten critical insights about how this feature works and what it means for publishers.

The core mechanism is an explicit-control layer that transforms your typed prompt into SEEMORE or SEELESS actions. These actions are applied after you confirm with a feed refresh. Behind the scenes, a large language model interprets your natural language input, converting it into structured instructions. The system maintains a persistent chat thread, meaning your tuning is an ongoing conversation rather than a one-off request. Your prompt becomes a set of standing instructions that influence content selection both in real time and over time. The feature’s back-end pipeline, identified as historicalnaturallanguagetuningcontent.f, functions as the historical twin of the live tuning pipeline, ensuring past prompts continue to shape your feed.

Content selection happens through two distinct mechanisms. The majority of cards come from entity and interest expansion, where your prompt is mapped to related topics and sources. This is why asking for one publisher often surfaces related content rather than just that specific site. A minority of cards, however, use a query-intent fan-out mechanism, where your prompt is decomposed into precise, natural-language retrieval queries. This is the same pattern seen in Generative Engine Optimization, where a single prompt generates multiple sub-queries that fetch content by semantic relevance, with no prerequisite for popularity.

The feature introduces visible attribution, including the “You asked to see” label on cards served by this pipeline. A prompt history appears in My Activity, and cards carry tags like “resulting from natural language tuning.” This transparency allows users to understand exactly which prompts influenced their feed.

Perhaps most significantly for publishers, this pipeline creates a popularity bypass. Classic Discover pipelines largely re-serve content that has already circulated widely. In contrast, “Tailor Your Feed” retrieves content based on relevance to the prompt, regardless of prior Discover distribution. Our tracking data shows that a majority of cards from this pipeline point to articles with no detectable prior Discover circulation. This includes niche sites and small creators: vegan recipe blogs, Mississippi Today, LinkedIn posts, niche Japanese-property guides, and other publishers outside the mainstream. Even VentureBeat appeared on a “niche sites” prompt, illustrating the retrieval’s behavior.

For publishers, this represents a genuine shift in how Discover visibility can be earned. Selection power moves to the user, who now explicitly demands content. This opens a third path to visibility for small, niche sites, alongside the traditional routes of implicit affinity and explicit follows. To capitalize, publishers should optimize for entities and topics (the dominant expansion mode) by being unambiguously about what users will name. They should also optimize for query-intent vocabulary (the fan-out mode) by phrasing titles, H1s, and intros to match the natural-language informational queries a prompt decomposes into.

However, publishers must remain clear-eyed. This is not a mass distribution channel; the pipeline shows essentially no growth over time, and Google promotes its cards cautiously. It serves the user who asked, not a broad audience. The feature is not publisher-triggerable; only the user can activate it. It remains EN-only and geographically limited to Search Labs in the US, with adoption still early. The strategic read is that if this feature graduates from Search Labs and users embrace prompt-based tuning at scale, the popularity-agnostic retrieval it relies on is structurally favorable to small, focused, well-described sites that classic, popularity-dominated pipelines rarely reward.

What comes next depends on the French and EU rollout, which remains uncertain. Adoption rates will determine whether this matters for publishers; the empty My Activity surface suggests prompt-based tuning is still a niche behavior. The current and historical pipeline pair suggests tuning is meant to last over time as a standing instruction. A nascent generativeretrieval.f pipeline spotted in tracking data hints that LLM-driven retrieval may extend beyond this single feature. The bigger picture is clear: Discover is moving from observed personalization toward declared personalization, and the retrieval that serves declared intent doesn’t lock onto popularity. That’s the structural opening for niche publishers, if and only if the feature ships broadly and users adopt it.