Introduction
Google is handing more editorial power to users with a prompt driven Discover control that lets people tell the app what they want to see. The feature, which began testing in Search Labs late in 2025 and has evolved through spring 2026, converts plain English instructions into actions that nudge the Discover feed toward or away from topics, creators and publishers. The clearest account of the tracking and mechanics comes from Damien Andell at Search Engine Land, whose field reporting and pipeline analysis lays out how the system behaves in practice. See the full report on Search Engine Land. https://searchengineland.com/tailor-your-feed-google-discover-niche-sites-480777
What it does – simple controls, consequential effects
At face value the interface is straightforward. A prompt field at the top of Discover invites users to say what they want to see. The UI presents starter chips such as Show me content from and Keep me updated, or accepts free text. Behind that simple prompt the app translates the user sentence into one or more actions – for example SEE_MORE or SEE_LESS – and applies those after a feed refresh. The result is a declared personalization model: rather than passively inferring interests from behaviour, Discover will now accept an explicit, written user preference and try to honour it.
How it works under the hood
Andell’s tracking shows a two part architecture. First there is an apparently LLM driven interpreter that turns the user sentence into human readable guidance plus structured instructions. That conversation is persistent – the tuning is stored as a thread – and the app surfaces a history of past prompts in My Activity. Second, a dedicated selection pipeline (identified in metadata as naturallanguagetuningcontent.f and its historical twin historicalnaturallanguagetuningcontent.f) takes those instructions and retrieves cards for the feed.
Two retrieval modes are visible. Most often the system does entity and interest expansion – map a prompt to entities and related topics and surface items from that expanded set. Less commonly it decomposes the prompt into a batch of specific natural language retrieval queries, a fan-out that fetches articles by semantic relevance rather than by prior popularity. That latter behaviour is the functional analogue of the GEO fan-out now familiar to people tracking generative retrieval.
A six month rollout, with clear milestones
The feature first appeared in Search Labs in December 2025 with a chat style entry and the four core intents: SEE_MORE, KEEP_UPDATED, CREATOR_MORE and SEE_LESS. By mid January 2026 attribution tags started appearing on cards marked resulting from natural language tuning and a prompt history showed in My Activity. Through February the system introduced a historical tag so cards could be labelled as influenced by past prompts, and testers noted that SEE_LESS is not a full erase – it reduces prominence rather than deleting a topic. In April 2026 the UI rebranded to Add topics to your feed and added a You asked to see label to identify pipeline-driven cards. By late May testers confirmed that a minority of cards are driven by the query-intent fan-out. Evidence and timelines are drawn from Andell’s field captures and pipeline tracing. https://searchengineland.com/tailor-your-feed-google-discover-niche-sites-480777
Why smaller, niche publishers may benefit
This is the most meaningful structural change for publishers. Classic Discover pipelines favour content that has already proven popular. The new natural language tuning pipeline is different: the retrieval reaches for semantically relevant content regardless of whether it previously circulated in Discover. In practice that produced a high share of cards pointing to sites with no detectable prior Discover distribution in the tracking dataset – independent creators, niche blogs and specialist outlets appeared where the traditional feed would usually surface mainstream publishers.
For publishers this creates a third path to Discover visibility. Historically a small site could only reach the feed either by being picked up through implicit signals of sustained user interest or via a follow. Now a user-initiated prompt can put a niche site directly in front of that same user. That matters because it changes the discovery economics: topical precision, clear entity signals and alignment with natural-language informational queries become concrete optimisation levers.
Practical takeaways for content teams
– Double down on topical clarity – be explicitly about what you cover so entity expansion is unambiguous.
– Write titles and introductions that map to natural language queries – the fan-out matches full phrases and informational formulations rather than single keywords.
– Expect a targeted, not mass, effect – the pipeline is cautious and does not produce the same viral distribution as classic news pipelines.
– You cannot trigger the feature directly; you can only be retrieval ready when a user asks.
Limits and potential pitfalls
There are important caveats. The feature is currently English only and limited to Search Labs in the US; adoption remains low in early tests, with My Activity often empty, so the aggregate impact is uncertain unless the tool rolls out widely and users actually choose to use it. The retrieval can be loose – testers recorded false positives where surface word overlap returned irrelevant articles – and Google appears to promote these cards cautiously and to pull them back when matches are weak. That cautious approach reduces the risk of spam, but it also reduces the commercial upside for publishers hoping for scale.
There are also regulatory and platform risk points to watch. If the feature expands into the EU it will intersect with a different set of rules and sensitivities around algorithmic transparency and personalised recommendations. Publishers and advertisers should be mindful that declared personalisation fragments the audience: a site may become highly visible to a narrow cohort while remaining invisible at scale.
What this means for advertisers and platform strategy
From an advertising perspective the feature fragments intent signals. Advertisers can benefit if targeting and measurement adjust to capture users who explicitly request niche topics, but monetisation depends on whether these tuned feeds carry the kind of engagement that supports ad yield. Platform operators and commercial teams should not assume demand translates automatically into spend. For platforms with subscription or membership models, declared personalisation could increase perceived value for power users, but only if the feature becomes broadly used.
What to watch next
Three developments will decide whether this matters at scale. First, will Google expand from Search Labs US into other languages and jurisdictions, especially France and the EU where the rollout is currently absent. Second, will user adoption move beyond early testers so the declared preferences meaningfully change distribution patterns. Third, will the engine that interprets prompts and drives retrieval mature so false positives fall and promotion policies loosen. Andell’s tracking also hints at a nascent generativeretrieval.f pipeline, which if confirmed would suggest the LLM driven retrieval model could spread beyond this single feature. For publishers and marketers the prudent course is to prepare for a world where declared intent and semantic retrieval matter, but to be realistic about the short term payoff.
Conclusion
Add topics to your feed is a modest UI change with potentially large structural consequences. It hands selection power to users, creates a demand driven path into Discover for small sites and forces publishers to think in full phrases and explicit topics rather than implicit popularity signals. Yet its current footprint is narrow, the system is cautious and adoption is early. For now it is a promising opening rather than a revolution. Publishers should treat it as a technical opportunity to be prepared for, not as a replacement for established distribution strategies.
Further reading
Damien Andell, Tailor Your Feed: How Google Discover surfacing niche sites via natural language tuning. Search Engine Land. https://searchengineland.com/tailor-your-feed-google-discover-niche-sites-480777
Google Discover help. https://support.google.com/websearch/answer/7021184
Google Search Labs. https://search.google.com/search/labs
