Tailor Your Feed: The Google Discover fan-out that surfaces niche sites

“Tailor Your Feed” is the first time a user can shape their Discover feed by typing, in natural language, what they want to see. We have tracked it from its first appearance in Search Labs^search-labs to the pipeline[^pipeline] that powers it. Ten key points:

1. An explicit-control layer. Your prompt is turned into SEE_MORE / SEE_LESS actions, applied after a feed refresh.
2. Seemingly an LLM[^llm] under the hood. A persistent chat thread, and your prompt turned into instructions applied to your feed (in real time and over time).
3. The rebrand. “Tailor Your Feed” became “Add topics to your feed” in spring 2026, with a chat-style entry point.
4. The back-end pipeline. historicalnaturallanguagetuningcontent.f[^pipeline-id], the “historical” twin of naturallanguagetuningcontent.f.
5. Two ways content is chosen. Entity[^entity] / interest expansion (the majority) vs a query-intent[^query-intent] fan-out[^fan-out] (the minority), the latter being the GEO[^geo] mechanism inside Discover.
6. Visible attribution. The “You asked to see” label, the “resulting from natural language tuning” tag, and a prompt history in My Activity.
7. Niche sites and small creators surfaced. Vegan recipe creators, Mississippi Today, a LinkedIn post, niche Japanese-property blogs and, as an illustration of the retrieval’s[^retrieval] behaviour, publishers outside the usual mainstream (VentureBeat surfaced on a “niche sites” prompt, though not itself a small site).
8. A popularity bypass. This pipeline mostly carries content that had barely circulated in Discover before, the opposite of the classic pipelines that re-serve already-popular articles.
9. What it changes for publishers. Selection power shifts to the user, opening a third path to visibility for small, niche sites.
10. Still EN-only, still nascent. Search Labs US only (FR ≈ 0%), adoption still early. What’s next.

Methodology

This article combines two observation streams:

• Field tracking of the feature in the Google app since December 2025: UI states, server responses, attribution tags, and feed behaviour after each “Refresh / Update your feed”. Captured on our test devices, US (English) Search Labs accounts.
• A close reading of the feed itself: each card can be traced back to the pipeline that selected it. By isolating the cards served by historicalnaturallanguagetuningcontent.f, we describe how this pipeline behaves relative to the rest of the feed, drawing on 1492.vision tracking data.

Three deliberate notes on how we phrase things:

• We describe distribution outcomes[^distribution], whether an article had ever circulated in Discover before, not raw audience numbers. When we say a card has “no prior Discover distribution”, we mean we find no trace of earlier serving in our Discover tracking dataset.
• No account identifier appears in this article. Examples are shown as prompt → result, anonymised.

The internal mechanisms below are our interpretation of observed data and public research. Where a date is inferred rather than anchored, we say so.

1. What “Tailor Your Feed” is: an explicit-control layer

For years, Discover personalization was implicit: Google inferred your interests from clicks, dwell time, follows. “Tailor Your Feed” adds the opposite, an explicit layer where you simply type what you want.

The current entry point at the top of the feed: “What do you want to see?” with an “Add topics to your feed” field.

The feature opens a chat-style panel. You can pick a suggested template or write freely.

The original “Tailor your feed” intro card: “Say what you want in your own words”, with a “Try now” button.

Suggested prompts: “Start showing me women’s basketball”, “Keep me updated on country music”, “Show me more of Cara Nicole’s videos”, with a free-text box “Ask for the kind of content you want.”

These suggestions correspond to four intents the server responses expose: SEE_MORE, KEEP_UPDATED, CREATOR_MORE and SEE_LESS. Whatever you type is interpreted into one (or several) of these actions, then applied to the feed once confirmed with “Refresh / Update your feed.”

The feature shipped through Search Labs, US only.

The Search Labs entry: “Make your Discover feed truly yours by saying what you want to see.” Beta, US only.

Later iterations turned the free-text box into explicit starter templates: “Show me content from…”, “I want videos about…”, “Keep me updated…”. The same verbs, now surfaced as chips.

The “What do you want to see?” panel with its starter templates: “Show me content from…”, “I want videos about…”, “Keep me updated…”.

2. Under the hood: seemingly an LLM that turns prompts into actions

The flow is simple: prompt → interpretation → readable answer + an actionable result. You type a prompt, the assistant replies in plain language and proposes concrete changes, and a tap on “Update your feed” commits them.

A representative response, observed in the data exchanges for the prompt “show me more content on seroundtable.com”, looks like this:

{
  "feature": "Discover • Tailor your feed",
  "locale": "en-US",
  "thread_key": "chat_thread_key_082fa565-234a-451c-9318-1e9af8b9d734",
  "user_prompt": "show me more content on seroundtable.com",
  "assistant_text": "I can show you more content from Search Engine Roundtable (SERoundtable.com) related to your interests. Refresh your feed to apply these changes.",
  "result": {
    "status": "UNDERSTOOD_AND_ACTIONABLE",
    "actions": ["SEE_MORE"],
    "show_call_to_action": true,
    "count": 1
  },
  "ui_state_code": 2
}

The in-app prompt that produced the response above: “show me more content on seroundtable.com”, and the assistant’s reply ending on “Refresh your feed.”

Three things stand out:

• A persistent thread. The thread_key is stable across exchanges; your tuning is a conversation, not a one-shot. The same thread is referenced again when, later, a card is attributed to one of your past prompts.
• Actions, not topics. The response returns actions: ["SEE_MORE"]. Ask to remove a topic and you get ["SEE_LESS"]; a nuanced prompt can return both, e.g. “new country music releases… but no celebrity gossip” yields ["SEE_MORE", "SEE_LESS"].
• Local context injection. Responses are interpreted with your context (locale, language, location). A generic “keep me updated about NBA” came back with “Updates on the Brooklyn Nets“, a local team injected from context.

“keep me updated about nba” → the assistant proposes scores, team and player news, and “Updates on the Brooklyn Nets”, a locally-injected entity.

On Google’s side, your sentence is turned into a set of instructions that feed the retrieval stage, with an “offline” path (applied over time) and a real-time one. This is, as we read it, the architectural shift the feature embodies: from inferred interest vectors to a natural-language profile you write yourself.

3. A six-month timeline (December 2025 → June 2026)

We documented the rollout as it happened. Dates anchored to in-app captures and the feature’s own changelogs; a few are approximate (marked with “~”).

December 2025: first sighting (Search Labs, US). The feature appears with everything described above: the chat panel, the JSON response, the persistent thread, the four intents, local context injection. First impression: the effect is subtle; after several refreshes you see a few on-topic cards, but nothing dramatic. (field note, example)

“I need a break from negative news. Show me more feel-good stories, but keep the local and breaking news.” The assistant lists the kinds of uplifting content to expect, then offers “Refresh your feed.”

Another early example: “can you show me https://dev.to/ on my feed” → the assistant offers programming articles, web-dev tutorials and software news.

~January 13, 2026: the attribution tag. Google starts marking cards “resulting from natural language tuning” after a refresh and the SEE_MORE/SEE_LESS arbitration, making it possible, for the first time, to tell which cards a prompt actually changed. A prompt history also appears in My Activity. (field note)

“‘Tailor your feed’ preferences in Discover. View and manage your prompts.” A dedicated My Activity surface (here, empty).

~February 2026: “historical” tuning, and what SEE_LESS really does. A second tag appears, “historical natural language tuning”, for cards influenced by a past prompt. Testing “fewer X posts”, we saw Google replace X (Twitter) cards with YouTube videos, and, notably, asking to remove a topic does not truly remove it: you get SEE_LESS, but the topic isn’t deleted from the feed. (field note)

~February 2026: the “niche” test. Asked for “more niche / small sites”, the feed came back, on the first refresh, with 2 of 10 cards modified by the prompt (a one-off snapshot, not an average), surfacing VentureBeat and Mississippi Today, with the very first result driven by the request. (field note)

~February 2026: the “entity” test. Asked for “more articles from a specific creator”, Google understood the topics related to that creator (entities), refreshed, and surfaced a LinkedIn post from them, tagged “natural language tuning content”. (field note)

~April 2026: the rebrand + the “You asked to see” label. “Tailor your feed” becomes “Add topics to your feed” with a chat UI, and a visible “You asked to see” label now marks the cards served by the pipeline. (rebrand, label)

A card labeled “You asked to see”: a historical-figures listicle from AOL.

“You asked to see” on a Guardian politics story.

“You asked to see” on Search Engine Roundtable cards; the publisher requested earlier now surfaces explicitly.

May 22, 2026: the query intents (the fan-out). We confirm that, beyond entity expansion, some pipeline cards carry a stored query intent, the prompt decomposed into specific retrieval queries, on the same pattern as the GEO “fan-out”. (More in section 5.) (field note)

“You asked to see” also applies to video: a Crunchyroll/One Piece YouTube card surfaced by a prompt.

June 2026: current state. The entry point reads “Add topics to your feed”, and the feature now surfaces small creators well outside the major publishers (section 6).

4. The pipeline behind it: historicalnaturallanguagetuningcontent.f

Every Discover card can be traced back to the pipeline that selected it. “Tailor Your Feed” maps to a dedicated pair:

• naturallanguagetuningcontent.f, content based on your current natural-language preferences.
• historicalnaturallanguagetuningcontent.f, content based on past prompts that keep influencing the feed (the “historical” tag from the timeline).

The pipeline retrieves content in two distinct ways:

• Mode A: entity / interest expansion (the majority). Based on the observed behaviour, the prompt is mapped to entities and topics, and the feed expands around them. This is why asking for one publisher surfaces related sources and topics, not just that publisher, the same logic as the Follow button. Most cards work this way, expanding around your topics rather than echoing the exact words you typed.
• Mode B: query-intent fan-out (the minority). For a fraction of cards, the prompt is decomposed into explicit query intents, natural-language retrieval queries that fetch the article. This is the GEO “fan-out” mechanism, and it is the subject of section 5.

One behaviour worth flagging: in our tracking data, Google appears to promote these cards cautiously, on average less than other pipelines, and pulls them back more often than any other, consistent with a retrieval that sometimes matches loosely (we’ll see a concrete false positive in section 5). It is, by design, a targeted pipeline, not a mass-distribution one: its cards show essentially no growth over time, the lowest of any pipeline. It serves what was asked for, to the user who asked. It doesn’t snowball.

5. Query intent: the GEO “fan-out” inside Discover

This is the most interesting mechanism. For a slice of cards, the prompt is broken down into a specific query intent that matched the article, the prompt turned into precise, natural-language retrieval queries. It is the functional analogue of the fan-out described for Generative Engine Optimization: a single prompt is decomposed into sub-queries that retrieve content by semantic relevance, with no popularity prerequisite.

The decomposition is visible. A prompt about SEO becomes a set of informational queries:

And those query intents then retrieve real articles. Here are anonymised query intent → URL pairs we observed (the formulations are the exact internal query intents):

Our analyzer groups these query intents into clusters:

Clusters of query intents observed in the pipeline: “learning Google algorithms”, “AI content machine learning SEO”, “buying Japanese property guide”, “healthy cooking techniques”, “anime recommendations 2026″…

The same clusters, with the article count per intent.

Drilling into one cluster shows both the strength and the limit of the fan-out:

The “buying Japanese property guide” cluster: japantoday.com (how to buy a home in Japan) and maigomika.com (rural Japan) are spot-on niche matches, but rockpapershotgun.com (Forza Horizon 6 in-game home locations) is a false positive, a video-game article pulled in by surface word overlap. Loose matches like this are why Google ranks this pipeline so cautiously (section 4).

Why this matters for SEO: the query intent reveals the exact vocabulary Google uses to map a prompt to your content. These are natural-language informational queries, not raw keywords. Aligning titles, H1s and intros with these formulations is the Discover-side equivalent of optimizing for the AI fan-out.

6. Niche sites and small creators: the popularity bypass

Classic Discover pipelines mostly re-serve content that is already popular, articles that have already circulated widely and built engagement. “Tailor Your Feed” works differently: the cards we observe show a retrieval that reaches for semantically relevant content regardless of whether it ever circulated in Discover before.

1492.vision tracking data backs this up. On historicalnaturallanguagetuningcontent.f, a majority of the cards point to articles with no detectable prior Discover distribution in our dataset, content that had never (or barely) been served in the feed before. This is, by a wide margin, the highest share of any pipeline: the classic news pipelines show the opposite profile, where almost every card has a long distribution history. A minority of the pipeline’s cards are the exception, mainstream articles (a Reuters story, for instance) already widely distributed and simply re-surfaced here.

The clearest illustration is a recipe prompt. Asking for vegan recipes surfaced, not the big food publishers, but small independent creators:

“Show me content from recipes vegan” → the assistant proposes plant-based weeknight dinners, vegan stews, high-protein tofu, vegan baking… then “Update your feed.”

The result: a “Sweet Potato Tacos” recipe from an independent creator, and “72 Vegan BBQ Recipes” from a small vegan blog, each labelled “You asked to see”, with a rating widget (“How would you rate this suggestion?”).

Across our tracking, the same pattern recurs: a niche gardening blog (buzzyseeds.com) for a seed-starting prompt; Mississippi Today for a “niche sites” prompt (with VentureBeat surfaced on the same prompt, as an illustration of the retrieval’s behaviour); a LinkedIn post for a creator prompt; niche Japanese-property blogs (japantoday.com, maigomika.com) for a property prompt. Most are not the usual high-volume Discover winners.

The takeaway, carefully put: the feature surfaces articles that had barely circulated in Discover before. The retrieval appears to be driven by relevance to the prompt, not by prior popularity: what works like a popularity filter in the classic pipelines is, here, bypassed.

7. What this changes for publishers

This is the part that matters most looking forward, and it’s a genuine shift in how Discover visibility can be earned.

Selection power moves to the user. In the classic feed, Google decides what you see from inferred signals. Here, the user writes the prompt (“show me more of X”, “less of Y”), and Google turns it into entities, interests and query intents that drive retrieval. Demand becomes explicit.

Consequence: niche sites can surface without a Discover track record. Because the retrieval appears to reach for relevance rather than prior popularity, a small site can be served the moment a user asks for its topic, even if it had never really circulated in Discover before (section 6). That’s new.

It’s a third path to visibility. Until now, a niche site broke into Discover only two ways: through strong implicit affinity (Google infers, from repeated engagement, that you love a topic, and a re-surface pipeline keeps feeding you that niche site), or through an explicit follow. “Tailor Your Feed” adds a third, user-initiated path that depends on neither.

The concrete levers for publishers:

• Optimize for entities/topics (the dominant Mode A). Be unambiguously about what users will name. A clear topical focus → cleaner entity association → you’re in the expansion set when someone asks for your subject.
• Optimize for query-intent vocabulary (the Mode B fan-out). Phrase titles, H1s and intros to match the natural-language informational queries a prompt decomposes into (the Discover-side of GEO). Section 5 shows the exact formulations Google uses.

What it is not. Publishers should be clear-eyed:

• Not a mass channel. The pipeline shows essentially no growth, and Google promotes its cards cautiously. It serves the user who asked; it doesn’t broadcast.
• Not publisher-triggerable. Only the user can fire it. You can be retrieval-ready, but you can’t activate it for your own site.
• Geographically and adoption-limited. It is EN-only (Search Labs, US; ≈ 0% in French feeds), and adoption is still early; the My Activity surface was empty in our tests. The future impact depends on (a) a general rollout and (b) whether users actually adopt prompt-based tuning at scale.

The strategic read: if “Add topics to your feed” graduates from Search Labs and users embrace it, the demand-driven, popularity-agnostic retrieval it relies on is structurally favourable to small, focused, well-described sites, the kind that classic, popularity-dominated pipelines rarely reward.

What’s next?

A snapshot, not an endpoint. What we’re watching:

• The French (and EU) rollout. Today the feature is EN-only. Depending on how independent it is from AIO[^aio] and AI Mode[^ai-mode] features, it could reach France sooner or later.
• Adoption. A powerful mechanism with no users changes nothing. The empty My Activity surface suggests prompt-based tuning is still a niche behaviour. Mass adoption is the variable that decides whether this matters for publishers.
• The current / historical pair. naturallanguagetuningcontent.f (live) and historicalnaturallanguagetuningcontent.f (persistent) suggest tuning is meant to last over time, a standing instruction, not a one-off.
• Convergence with generative retrieval. A nascent generativeretrieval.f pipeline, spotted recently in our tracking data, suggests LLM-driven retrieval may reach beyond this one feature (to be confirmed).

The bigger picture: Discover is moving from observed personalization (Google infers) toward declared personalization (you tell it), 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.

Notes

[^pipeline]: In Discover, a pipeline is a selection circuit that chooses and serves the cards. Each card can be traced to the identifier of the pipeline that produced it; that is what we exploit in our 1492.vision tracking data.

[^pipeline-id]: The .f suffix in identifiers such as historicalnaturallanguagetuningcontent.f is an internal marker we observe in the metadata of Discover cards; it distinguishes the selection circuits.

[^search-labs]: Google’s beta program that lets users test experimental features in Search and Discover before any wider rollout.

[^llm]: Large Language Model. We assume one here, without direct proof of the model used, from the conversational behaviour and the structured responses observed.

[^entity]: In the sense of Google’s Knowledge Graph: a named entity; a person, topic, organisation or concept that is identified and linked to others, distinct from the everyday sense of the word “entity” (a company, a legal structure).

[^fan-out]: A mechanism by which a single prompt is broken into several retrieval sub-queries, each targeting a different angle of the topic.

[^query-intent]: A natural-language informational query, derived from the user’s prompt, that was used to retrieve a specific article. Observable in our data as metadata attached to certain pipeline cards.

[^geo]: Generative Engine Optimization: the practice of optimizing content to be visible and cited in the answers of generative engines (AI Overviews, ChatGPT, etc.), notably via query fan-out.

[^retrieval]: The stage where Google fetches and selects the content to display in the feed, from interest signals, entities or queries.

[^distribution]: In this article, the fact that an article had already circulated in Discover, as observed in our 1492.vision tracking dataset, not an audience figure nor a Search Console metric.

[^aio]: AI Overviews: generative answers shown at the top of certain Google Search results.

[^ai-mode]: AI Mode: Google Search’s conversational interface, distinct from AI Overviews.

Field tracking: Google app, Search Labs (US/English) accounts, December 2025 – June 2026. Pipeline behaviour derived from close observation of the Discover feed via 1492.vision tracking data. “Distribution” here means whether an article had already circulated in Discover, as observed in our tracking dataset, not private audience figures. The internal mechanisms are our interpretation of observed data and public research; inferred dates are flagged as approximate.