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

Jun 24, 2026 07:00 PM - 4 hours ago 106

“Tailor Your Feed” is the first clip a personification tin style their Discover provender by typing, successful earthy language, what they want to see. We person tracked it from its first quality successful Search Labs^search-labs to the pipeline[^pipeline] that powers it. Ten cardinal points:

An explicit-control layer. Your punctual is turned into SEE_MORE / SEE_LESS actions, applied aft a provender refresh.
Seemingly an LLM[^llm] nether the hood. A persistent chat thread, and your punctual turned into instructions applied to your provender (in existent clip and complete time).
The rebrand. “Tailor Your Feed” became “Add topics to your feed” successful outpouring 2026, pinch a chat-style introduction point.
The back-end pipeline. historicalnaturallanguagetuningcontent.f[^pipeline-id], the “historical” copy of naturallanguagetuningcontent.f.
Two ways contented is chosen. Entity[^entity] / liking description (the majority) vs a query-intent[^query-intent] fan-out[^fan-out] (the minority), the second being the GEO[^geo] system wrong Discover.
Visible attribution. The “You asked to see” label, the “resulting from earthy connection tuning” tag, and a punctual history successful My Activity.
Niche sites and mini creators surfaced. Vegan look creators, Mississippi Today, a LinkedIn post, niche Japanese-property blogs and, arsenic an illustration of the retrieval’s[^retrieval] behaviour, publishers extracurricular the accustomed mainstream (VentureBeat surfaced connected a “niche sites” prompt, though not itself a mini site).
A fame bypass. This pipeline mostly carries contented that had barely circulated successful Discover before, the other of the classical pipelines that re-serve already-popular articles.
What it changes for publishers. Selection powerfulness shifts to the user, opening a 3rd way to visibility for small, niche sites.
Still EN-only, still nascent. Search Labs US only (FR ≈ 0%), take still early. What’s next.

Methodology

This article combines 2 study streams:

Field tracking of the characteristic successful the Google app since December 2025: UI states, server responses, attribution tags, and provender behaviour aft each “Refresh / Update your feed”. Captured connected our trial devices, US (English) Search Labs accounts.
A adjacent reference of the provender itself: each paper tin beryllium traced backmost to the pipeline that selected it. By isolating the cards served by historicalnaturallanguagetuningcontent.f, we picture really this pipeline behaves comparative to the remainder of the feed, drafting connected 1492.vision search data.

Three deliberate notes connected really we building things:

We picture distribution outcomes[^distribution], whether an article had ever circulated successful Discover before, not earthy assemblage numbers. When we opportunity a paper has “no anterior Discover distribution”, we mean we find nary trace of earlier serving successful our Discover search dataset.
No relationship identifier appears successful this article. Examples are shown arsenic punctual → result, anonymised.

The soul mechanisms beneath are our mentation of observed information and nationalist research. Where a day is inferred alternatively than anchored, we opportunity 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 furniture wherever you simply type what you want.

The existent introduction constituent astatine the apical of the feed: “What do you want to see?” pinch an “Add topics to your feed” field.

The characteristic opens a chat-style panel. You tin prime a suggested template aliases constitute freely.

The original “Tailor your feed” intro card: “Say what you want successful your ain words”, pinch a “Try now” button.

Suggested prompts: “Start showing maine women’s basketball”, “Keep maine updated connected state music”, “Show maine much of Cara Nicole’s videos”, pinch a free-text container “Ask for the benignant of contented you want.”

These suggestions correspond to 4 intents the server responses expose: SEE_MORE, KEEP_UPDATED, CREATOR_MORE and SEE_LESS. Whatever you type is interpreted into 1 (or several) of these actions, past applied to the provender erstwhile confirmed pinch “Refresh / Update your feed.”

The characteristic shipped done Search Labs, US only.

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

Later iterations turned the free-text container into definitive starter templates: “Show maine contented from…”, “I want videos about…”, “Keep maine updated…”. The aforesaid verbs, now surfaced arsenic chips.

The “What do you want to see?” sheet pinch its starter templates: “Show maine contented from…”, “I want videos about…”, “Keep maine updated…”.

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

The travel is simple: prompt → mentation → readable reply + an actionable result. You type a prompt, the adjunct replies successful plain connection and proposes actual changes, and a pat connected “Update your feed” commits them.

A typical response, observed successful the information exchanges for the punctual “show maine much contented connected seroundtable.com”, looks for illustration this:

{ "feature": "Discover • Tailor your feed", "locale": "en-US", "thread_key": "chat_thread_key_082fa565-234a-451c-9318-1e9af8b9d734", "user_prompt": "show maine much contented connected seroundtable.com", "assistant_text": "I tin show you much contented from MCP (SERoundtable.com) related to your interests. Refresh your provender to use these changes.", "result": { "status": "UNDERSTOOD_AND_ACTIONABLE", "actions": ["SEE_MORE"], "show_call_to_action": true, "count": 1 }, "ui_state_code": 2 }

The in-app punctual that produced the consequence above: “show maine much contented connected seroundtable.com”, and the assistant’s reply ending connected “Refresh your feed.”

Three things guidelines out:

A persistent thread. The thread_key is unchangeable crossed exchanges; your tuning is simply a conversation, not a one-shot. The aforesaid thread is referenced again when, later, a paper is attributed to 1 of your past prompts.
Actions, not topics. The consequence returns actions: ["SEE_MORE"]. Ask to remove a taxable and you get ["SEE_LESS"]; a nuanced punctual tin return both, e.g. “new state euphony releases… but nary personage gossip” yields ["SEE_MORE", "SEE_LESS"].
Local discourse injection. Responses are interpreted pinch your discourse (locale, language, location). A generic “keep maine updated astir NBA” came backmost pinch “Updates connected the Brooklyn Nets“, a section squad injected from context.

“keep maine updated astir nba” → the adjunct proposes scores, squad and subordinate news, and “Updates connected the Brooklyn Nets”, a locally-injected entity.

On Google’s side, your condemnation is turned into a group of instructions that provender the retrieval stage, pinch an “offline” way (applied complete time) and a real-time one. This is, arsenic we publication it, the architectural displacement the characteristic embodies: from inferred liking vectors to a natural-language floor plan you constitute yourself.

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

We documented the rollout arsenic it happened. Dates anchored to in-app captures and the feature’s ain changelogs; a fewer are approximate (marked pinch “~”).

December 2025: first sighting (Search Labs, US). The characteristic appears pinch everything described above: the chat panel, the JSON response, the persistent thread, the 4 intents, section discourse injection. First impression: the effect is subtle; aft respective refreshes you spot a fewer on-topic cards, but thing dramatic. (field note, example)

“I request a break from antagonistic news. Show maine much feel-good stories, but support the section and breaking news.” The adjunct lists the kinds of uplifting contented to expect, past offers “Refresh your feed.”

Another early example: “can you show maine https://dev.to/ connected my feed” → the adjunct offers programming articles, web-dev tutorials and package news.

~January 13, 2026: the attribution tag. Google starts marking cards “resulting from earthy connection tuning” aft a refresh and the SEE_MORE/SEE_LESS arbitration, making it possible, for the first time, to show which cards a punctual really changed. A prompt history besides appears successful My Activity. (field note)

“‘Tailor your feed’ preferences successful Discover. View and negociate your prompts.” A dedicated My Activity aboveground (here, empty).

~February 2026: “historical” tuning, and what SEE_LESS really does. A 2nd tag appears, “historical earthy connection tuning”, for cards influenced by a past prompt. Testing “fewer X posts”, we saw Google replace X (Twitter) cards pinch YouTube videos, and, notably, asking to remove a taxable does not genuinely region it: you get SEE_LESS, but the taxable isn’t deleted from the feed. (field note)

~February 2026: the “niche” test. Asked for “more niche / mini sites”, the provender came back, connected the first refresh, pinch 2 of 10 cards modified by the punctual (a one-off snapshot, not an average), surfacing VentureBeat and Mississippi Today, pinch the very first consequence driven by the request. (field note)

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

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

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

“You asked to see” connected a Guardian authorities story.

“You asked to see” connected MCP cards; the patient requested earlier now surfaces explicitly.

May 22, 2026: the query intents (the fan-out). We corroborate that, beyond entity description , some pipeline cards transportation a stored query intent, the punctual decomposed into circumstantial retrieval queries, connected the aforesaid shape arsenic the GEO “fan-out”. (More successful conception 5.) (field note)

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

June 2026: existent state. The introduction constituent sounds “Add topics to your feed”, and the characteristic now surfaces mini creators good extracurricular the awesome publishers (section 6).

4. The pipeline down it: historicalnaturallanguagetuningcontent.f

Every Discover paper tin beryllium traced backmost to the pipeline that selected it. “Tailor Your Feed” maps to a dedicated pair:

naturallanguagetuningcontent.f, contented based connected your current natural-language preferences.
historicalnaturallanguagetuningcontent.f, contented based connected past prompts that support influencing the provender (the “historical” tag from the timeline).

The pipeline retrieves contented successful two chopped ways:

Mode A: entity / liking description (the majority). Based connected the observed behaviour, the punctual is mapped to entities and topics, and the provender expands astir them. This is why asking for 1 patient surfaces related sources and topics, not conscionable that publisher, the aforesaid logic arsenic the Follow button. Most cards activity this way, expanding astir your topics alternatively than echoing the nonstop words you typed.
Mode B: query-intent fan-out (the minority). For a fraction of cards, the punctual is decomposed into definitive query intents, natural-language retrieval queries that fetch the article. This is the GEO “fan-out” mechanism, and it is the taxable of conception 5.

One behaviour worthy flagging: successful our search data, Google appears to beforehand these cards cautiously, connected mean little than different pipelines, and pulls them backmost much often than immoderate other, accordant pinch a retrieval that sometimes matches loosely (we’ll spot a actual mendacious affirmative successful conception 5). It is, by design, a targeted pipeline, not a mass-distribution one: its cards show fundamentally no growth complete time, the lowest of immoderate pipeline. It serves what was asked for, to the personification who asked. It doesn’t snowball.

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

This is the astir absorbing mechanism. For a portion of cards, the punctual is surgery down into a circumstantial query intent that matched the article, the punctual turned into precise, natural-language retrieval queries. It is the functional analogue of the fan-out described for Generative Engine Optimization: a azygous punctual is decomposed into sub-queries that retrieve contented by semantic relevance, pinch nary fame prerequisite.

The decomposition is visible. A punctual astir SEO becomes a group of informational queries:

User punctual (approx.)Decomposed query intents

“Show maine contented from SEO”	“SEO strategies algorithm changes” · “Google ranking strategy updates” · “tips for getting contented into google discover”

And those query intents past retrieve existent articles. Here are anonymised query intent → URL pairs we observed (the formulations are the nonstop soul query intents):

Query intentRetrieved articleProfile

starting seeds indoors guide	buzzyseeds.com/…/how-to-grow-strawberry-from-seed-indoors	niche gardening, nary anterior Discover distribution
buying Japanese spot guide	japantoday.com/…/how-to-buy-a-home-in-japan-as-a-foreigner	niche
buying Japanese spot guide	maigomika.com/…/rural-japan-inaka-levels	niche
personal stories surviving successful France	perfectlyprovence.co/…	niche
tips for getting contented into google discover	conductor.com/academy/best-aeo-geo-tools	mid-size
AI contented instrumentality learning SEO	searchengineland.com/ai-increase-seo-expertise-value	trade press
best sci-fi books	collider.com/best-sci-fi-books-last-25-years-ranked	mainstream
Nvidia banal analysis	reuters.com/technology/nvidia-invests-2-billion	mainstream, already wide distributed, simply re-surfaced

Our researcher groups these query intents into clusters:

Clusters of query intents observed successful the pipeline: “learning Google algorithms”, “AI contented instrumentality learning SEO”, “buying Japanese spot guide”, “healthy cooking techniques”, “anime recommendations 2026″…

The aforesaid clusters, pinch the article count per intent.

Drilling into 1 cluster shows some the spot and the limit of the fan-out:

The “buying Japanese spot guide” cluster: japantoday.com (how to bargain a location successful Japan) and maigomika.com (rural Japan) are spot-on niche matches, but rockpapershotgun.com (Forza Horizon 6 in-game location locations) is simply a false positive, a video-game article pulled successful by aboveground connection overlap. Loose matches for illustration this are why Google ranks this pipeline truthful cautiously (section 4).

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

6. Niche sites and mini creators: the fame bypass

Classic Discover pipelines mostly re-serve contented that is already popular, articles that person already circulated wide and built engagement. “Tailor Your Feed” useful differently: the cards we observe show a retrieval that reaches for semantically applicable contented regardless of whether it ever circulated successful Discover before.

1492.vision search information backs this up. On historicalnaturallanguagetuningcontent.f, a mostly of the cards constituent to articles pinch nary detectable anterior Discover distribution successful our dataset, contented that had ne'er (or barely) been served successful the provender before. This is, by a wide margin, the highest stock of immoderate pipeline: the classical news pipelines show the other profile, wherever almost each paper has a agelong distribution history. A number of the pipeline’s cards are the exception, mainstream articles (a Reuters story, for instance) already wide distributed and simply re-surfaced here.

The clearest illustration is simply a look prompt. Asking for vegan recipes surfaced, not the large nutrient publishers, but small independent creators:

“Show maine contented from recipes vegan” → the adjunct proposes plant-based weeknight dinners, vegan stews, high-protein tofu, vegan baking… past “Update your feed.”

The result: a “Sweet Potato Tacos” look from an independent creator, and “72 Vegan BBQ Recipes” from a mini vegan blog, each labelled “You asked to see”, pinch a standing widget (“How would you complaint this suggestion?”).

Across our tracking, the aforesaid shape recurs: a niche farming blog (buzzyseeds.com) for a seed-starting prompt; Mississippi Today for a “niche sites” punctual (with VentureBeat surfaced connected the aforesaid prompt, arsenic an illustration of the retrieval’s behaviour); a LinkedIn post for a creator prompt; niche Japanese-property blogs (japantoday.com, maigomika.com) for a spot prompt. Most are not the accustomed high-volume Discover winners.

The takeaway, cautiously put: the characteristic surfaces articles that had hardly circulated successful Discover before. The retrieval appears to beryllium driven by relevance to the prompt, not by anterior popularity: what useful for illustration a fame select successful the classical pipelines is, here, bypassed.

7. What this changes for publishers

This is the portion that matters astir looking forward, and it’s a genuine displacement successful really Discover visibility tin beryllium earned.

Selection powerfulness moves to the user. In the classical feed, Google decides what you spot from inferred signals. Here, the personification writes the prompt (“show maine much of X”, “less of Y”), and Google turns it into entities, interests and query intents that thrust retrieval. Demand becomes explicit.

Consequence: niche sites tin aboveground without a Discover way record. Because the retrieval appears to scope for relevance alternatively than anterior popularity, a mini tract tin beryllium served the infinitesimal a personification asks for its topic, moreover if it had never really circulated successful Discover before (section 6). That’s new.

It’s a 3rd way to visibility. Until now, a niche tract collapsed into Discover only 2 ways: done beardown implicit affinity (Google infers, from repeated engagement, that you emotion a topic, and a re-surface pipeline keeps feeding you that niche site), aliases done an definitive follow. “Tailor Your Feed” adds a third, user-initiated way that depends connected neither.

The actual levers for publishers:

Optimize for entities/topics (the ascendant Mode A). Be unambiguously about what users will name. A clear topical attraction → cleaner entity relation → you’re successful the description group erstwhile personification asks for your subject.
Optimize for query-intent vocabulary (the Mode B fan-out). Phrase titles, H1s and intros to lucifer the natural-language informational queries a punctual decomposes into (the Discover-side of GEO). Section 5 shows the nonstop formulations Google uses.

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

Not a wide channel. The pipeline shows fundamentally nary growth, and Google promotes its cards cautiously. It serves the personification who asked; it doesn’t broadcast.
Not publisher-triggerable. Only the personification tin occurrence it. You tin beryllium retrieval-ready, but you can’t activate it for your ain site.
Geographically and adoption-limited. It is EN-only (Search Labs, US; ≈ 0% successful French feeds), and take is still early; the My Activity aboveground was quiet successful our tests. The early effect depends connected (a) a wide rollout and (b) whether users really adopt prompt-based tuning astatine scale.

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

What’s next?

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

The French (and EU) rollout. Today the characteristic is EN-only. Depending connected really independent it is from AIO[^aio] and AI Mode[^ai-mode] features, it could scope France sooner aliases later.
Adoption. A powerful system pinch nary users changes nothing. The quiet My Activity aboveground suggests prompt-based tuning is still a niche behaviour. Mass take is the adaptable that decides whether this matters for publishers.
The existent / humanities pair. naturallanguagetuningcontent.f (live) and historicalnaturallanguagetuningcontent.f (persistent) propose tuning is meant to last complete time, a opinionated instruction, not a one-off.
Convergence pinch generative retrieval. A nascent generativeretrieval.f pipeline, spotted precocious successful our search data, suggests LLM-driven retrieval whitethorn scope beyond this 1 characteristic (to beryllium confirmed).

The bigger picture: Discover is moving from observed personalization (Google infers) toward declared personalization (you show it), and the retrieval that serves declared intent doesn’t fastener onto popularity. That’s the structural opening for niche publishers, if and only if the characteristic ships broadly and users adopt it.

Notes

[^pipeline]: In Discover, a pipeline is simply a action circuit that chooses and serves the cards. Each paper tin beryllium traced to the identifier of the pipeline that produced it; that is what we utilization successful our 1492.vision search data.

[^pipeline-id]: The .f suffix successful identifiers specified arsenic historicalnaturallanguagetuningcontent.f is an soul marker we observe successful the metadata of Discover cards; it distinguishes the action circuits.

[^search-labs]: Google’s beta programme that lets users trial experimental features successful Search and Discover earlier immoderate wider rollout.

[^llm]: Large Language Model. We presume 1 here, without nonstop impervious of the exemplary used, from the conversational behaviour and the system responses observed.

[^entity]: In the consciousness of Google’s Knowledge Graph: a named entity; a person, topic, organisation aliases conception that is identified and linked to others, chopped from the mundane consciousness of the connection “entity” (a company, a ineligible structure).

[^fan-out]: A system by which a azygous punctual is surgery into respective retrieval sub-queries, each targeting a different perspective of the topic.

[^query-intent]: A natural-language informational query, derived from the user’s prompt, that was utilized to retrieve a circumstantial article. Observable successful our information arsenic metadata attached to definite pipeline cards.

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

[^retrieval]: The shape wherever Google fetches and selects the contented to show successful the feed, from liking signals, entities aliases queries.

[^distribution]: In this article, the truth that an article had already circulated successful Discover, arsenic observed successful our 1492.vision search dataset, not an assemblage fig nor a Search Console metric.

[^aio]: AI Overviews: generative answers shown astatine the apical of definite Google Search results.

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

Field tracking: Google app, Search Labs (US/English) accounts, December 2025 – June 2026. Pipeline behaviour derived from adjacent study of the Discover provender via 1492.vision search data. “Distribution” present intends whether an article had already circulated successful Discover, arsenic observed successful our search dataset, not backstage assemblage figures. The soul mechanisms are our mentation of observed information and nationalist research; inferred dates are flagged arsenic approximate.