Google Discover’s New Pipeline Rewards Niche Publishers

TL;DR: Google’s “Tailor Your Feed” pipeline (now branded “Add topics to your feed”) uses LLM-interpreted prompts to serve niche content that has never circulated widely in Discover. The retrieval mechanism bypasses popularity as a filter, selecting by semantic relevance to the user’s declared intent. For operators building content in focused verticals, this is the first Discover path that doesn’t require an existing audience to get in front of a new one.

What “Tailor Your Feed” Actually Is

For years, Google Discover personalization was entirely implicit. Google watched what you clicked, how long you read, which topics you followed, and inferred your interests from those signals. The new feature breaks that pattern. A user opens the Discover feed, types a natural-language prompt into a chat panel, and Google interprets that prompt into one or more actions: SEE_MORE, SEE_LESS, KEEP_UPDATED, or CREATOR_MORE. Tap “Update your feed” and those instructions are applied, both immediately and persistently over time via a dedicated pipeline tagged historicalnaturallanguagetuningcontent.f.

The feature shipped through Google Search Labs in December 2025, US English only. It rebranded to “Add topics to your feed” in April 2026 with a cleaner chat-style UI and starter template chips. What did not change is the underlying mechanism: an apparent LLM interprets your sentence, replies in plain language, and converts your intent into structured retrieval instructions.

One observed server response for the prompt “show me more content on seroundtable.com” returned a status of UNDERSTOOD_AND_ACTIONABLE with actions: [“SEE_MORE”]. A nuanced prompt like “new country music releases but no celebrity gossip” returned both SEE_MORE and SEE_LESS simultaneously. The system also injects local context: “keep me updated about NBA” returned a proposal including “Updates on the Brooklyn Nets,” a locally-relevant entity pulled from the user’s location signal.

The Two Retrieval Modes Inside the Pipeline

The pipeline retrieves content in two distinct ways, and understanding the difference determines how you optimize for it.

Mode A: Entity and interest expansion (the majority of cards). The prompt is mapped to Knowledge Graph entities and related topics. Ask for one publisher and you get that publisher’s subject matter, not just that publisher. This is the same logic as the Follow button: topical expansion, not exact-match retrieval. If your site has a clear, unambiguous topical focus, you land in the expansion set when a user asks for your subject. Vague, multi-topic sites do not.

Mode B: Query-intent fan-out (a minority of cards). The prompt is decomposed into specific natural-language retrieval queries that fetch articles by semantic relevance. A prompt about SEO became query intents like “SEO strategies algorithm changes,” “Google ranking system updates,” and “tips for getting content into google discover.” Each of those intents then retrieved a real article. Observed matches included a niche gardening blog for a seed-starting prompt, a Japanese property blog for a buying-in-Japan prompt, and a small personal site about rural Japan, none of them mainstream publishers, none of them with prior Discover distribution in the tracking dataset.

This fan-out mechanism is structurally identical to what researchers describe in Generative Engine Optimization: a single input broken into sub-queries that retrieve by relevance, with no popularity prerequisite. The vocabulary of your title, H1, and opening paragraph determines whether a decomposed query intent maps to your article.

Why Niche Content Gets Surfaced Here

Classic Discover pipelines re-serve articles that already performed. An article needs Discover history to get more Discover distribution. That loop systematically excludes new or low-volume publishers.

The historicalnaturallanguagetuningcontent.f pipeline works the opposite way. Tracking data from 1492.vision showed that a majority of cards served by this pipeline pointed to articles with no detectable prior Discover distribution. That share is the highest of any pipeline observed. The retrieval reaches for semantic relevance to the prompt, not for prior engagement history. A small vegan recipe blog surfaced for a vegan cooking prompt. Mississippi Today surfaced for a “niche sites” prompt. A LinkedIn post from a specific creator surfaced after a user asked to see more from that creator.

This is not a rounding error. It is the structural difference between this pipeline and every other one: popularity is not the filter. Declared user intent is.

That said, the pipeline operates cautiously. Google promotes these cards less aggressively than other pipelines and pulls them back more often, consistent with a retrieval that sometimes matches loosely. One “buying Japanese property” cluster also surfaced a Forza Horizon 6 article about in-game home locations, a false positive from surface word overlap. Loose matches are the reason the pipeline does not snowball: it serves what was asked for, to the user who asked, and does not broadcast.

What This Means for Performance Marketing Operators

Operators running content programs in high-CAC verticals, whether that is iGaming acquisition, Forex lead generation, crypto user acquisition, or law firm intake marketing, have faced the same Discover wall as every other niche publisher: without existing distribution history, you rarely appear. This pipeline changes that equation, conditionally.

The conditions are straightforward. Your content needs a clear, narrow topical focus that maps cleanly to what a user would type in natural language. “Best online casino bonuses UK” is a query intent. A page that covers casino bonuses, poker strategy, sports betting tips, and financial news in one domain is not. Entity association requires topical clarity. Mode A expansion only works if Google can map your site to a specific, named topic or entity without ambiguity.

For Mode B, the optimization is more precise: phrase your titles and H1s to match how a user would decompose a high-intent question into sub-queries. Not keyword-stuffed anchor text, but the natural-language informational query that sits behind a user’s prompt. “How to open a Forex trading account in the US” is a query intent. “Forex trading guide” is a topic tag. These are not the same thing in retrieval terms.

Operators should also consider running a full content audit against their existing article inventory to identify which pieces have the topical specificity to qualify for Mode B fan-out and which are too broad to survive entity expansion. Content that was written for SEO keyword coverage rather than topical depth will not perform well here.

The precision targeting logic that works in paid channels applies directly to organic Discover optimization: narrow audience, specific intent, one clear topic per asset. Broad coverage pages do not get surfaced by a pipeline that is resolving a specific user prompt.

Attribution, Visibility Labels, and What Operators Can Track

One practical advantage of this pipeline is its transparency. Cards served by the natural language tuning pipeline carry visible labels: “resulting from natural language tuning” and “You asked to see” appear on the card face. A prompt history is logged in the user’s My Activity panel. This means users can see exactly which of their prompts triggered a piece of content.

For publishers, this creates a new content signal to watch: Search Console impressions from Discover that coincide with specific topical prompts. While Google does not expose the pipeline identifier in Search Console, an operator running paid and organic performance tracking in parallel can correlate Discover impression spikes with topical clusters and identify which content formats and title structures are landing in Mode B retrieval.

The attribution logic also matters for understanding what this pipeline is not. It is not a broadcast mechanism. It does not compound. A card served here shows essentially no growth over time in the tracking data. It serves the user who asked, once, and persists as a standing instruction. If 50,000 users ask to see content about a specific trading strategy, 50,000 individual pipeline instances fire, each one independent. Scale comes from the number of users who type a relevant prompt, not from viral re-serving.

Current Limitations and What to Watch

The feature is English-only and restricted to Search Labs US accounts as of June 2026. French-language feeds showed approximately zero penetration in the tracking data. Adoption is early: the My Activity surface was empty in test accounts, which suggests most users have not interacted with prompt-based tuning at all.

The two conditions that determine whether this matters at scale: a general rollout beyond Search Labs, and actual user adoption of explicit prompt-based feed tuning. If users don’t type prompts, the pipeline doesn’t fire. Right now, it is a mechanism with real structural implications that is operating at minimal volume.

Also visible in the tracking data is a nascent generativeretrieval.f pipeline, suggesting LLM-driven retrieval may extend beyond the Tailor Your Feed feature. That pipeline is unconfirmed in behavior, but its presence indicates Google is investing in prompt-driven retrieval as a broader Discover architecture, not a one-off experiment. Operators who want to position themselves ahead of that shift can work with AI-driven content qualification tools to align asset libraries with the natural-language query vocabulary that generative retrieval systems use.

The structural read is this: if “Add topics to your feed” ships broadly and users adopt it, the pipeline rewards focused, well-described, topically specific sites. That describes the content strategy that high-CAC operators in regulated verticals should already be running. The window to be retrieval-ready before mass adoption is now.

Originally reported by Search Engine Land, June 2026.