Key takeaways
- Custom Feeds let a viewer describe the feed they want in plain language and keep it as a saved tab on Home. YouTube announced the full version at Made on YouTube on 23 September 2026, for US viewers on web, mobile and TVs.
- The version already shipped is narrower than the headlines: YouTube's own Help Center calls it experimental, English-only and US-only, needs watch and search history switched on, and lets a prompt expire after 30 days of inactivity.
- YouTube has not published how videos are selected for a custom feed, and there is no creator-side control, no toggle and no new analytics line. Anyone selling you "custom feed optimisation" is guessing.
- The mechanism that matters is the shift from revealed preference to stated preference — from "people like you watched this" to "does this match the sentence I wrote". That changes which thumbnails win, not whether thumbnails matter.
- Inside a feed built from a sentence, fit beats mystery. The viewer already has intent; the image's job is to prove match faster than the nineteen near-identical videos next to it.
- Views from it land in the Browse features bucket with everything else from Home, so your home-feed CTR is about to average two populations with very different intent.
For most of YouTube's history the home page has been a single object. One feed, assembled for you, with no handle on it. You could subscribe, you could hit "not interested", you could clear your history and start again — but you could not tell it what you wanted in words. Every piece of packaging advice written since 2012 assumes that shape: a viewer arrives with no stated intent, and your thumbnail has to manufacture one.
That assumption now has an exception. YouTube is turning its prompt-built feed experiment into a shipped feature called Custom Feeds, where a viewer writes a sentence describing what they are in the mood for and gets a saved tab that keeps refreshing against it. The announcement landed at Made on YouTube on 23 September 2026 alongside thirty-odd other things, and it was the one viewer-facing change in that list that alters how a video gets found rather than how it gets made.
It is also the change most likely to be oversold to you in the next six months. So this piece does two jobs: establish precisely what exists today and what is still a press release, and then work out what genuinely follows for packaging — which turns out to be less about new tricks than about which of the old ones stop working when the viewer has already said what they want.
What is actually live, and what was announced
These are two different things, and most coverage blurs them. The feature has been arriving in stages for nearly a year.
| When | What happened | Status |
|---|---|---|
| November 2025 | YouTube began testing a way to reshape Home recommendations with a short prompt, surfaced as a chip next to "Home" for people in the test | Limited experiment |
| Mid-2026 | Reporting describes a broader US experiment: a "Your custom feed" chip at the top of Home, one active feed per account, prompt editable at any time | Experiment, eligible accounts only |
| 23 September 2026 | Announced as Custom Feeds at Made on YouTube: multiple saved feeds, for US viewers on web, mobile and TVs, described as arriving this autumn | Announced; rollout in progress |
YouTube's own Help Center page for the shipped version, "Create & manage your custom feed on YouTube Home", is the document to trust over any roundup. It describes the feature as experimental and available in English in the United States. You sign in, go to Home, open the custom feed tab, and either write a prompt or pick a suggested one — the example YouTube itself uses is "Help me unwind after work with guided meditations under 10 minutes". You can rewrite the prompt whenever you like. There is a "Something wrong?" option for when the feed misreads you. And a prompt expires after 30 days of inactivity, which is the single most revealing detail on the page.
Beyond that: watch history and search history have to be on, because the feed blends your sentence with your behaviour rather than replacing it. Reporting on the multi-feed test says device language must be English and the account must be in the US, and puts the cap for the multi-feed version at eight active feeds per account. On mobile, the custom feed tab takes the place of the familiar topic chips along the top of Home; on desktop, the classic chips are still there. Coverage consistently describes Gemini as the component interpreting the prompt. And the main Home feed does not go away — Custom Feeds sit beside it, not instead of it.
The detail nobody is pricing in
Prompts expire after 30 days of inactivity. A feature whose state decays unless re-used is a feature its own designers expect most people to try once. That is not a criticism — it is a sensible default — but it tells you how to size the opportunity. This is a surface that will matter enormously to a minority of viewers and not at all to most, and the right amount of channel strategy to rebuild around it is none.
Two different questions a feed can ask about your video
To work out what changes for packaging you have to be precise about what changes for retrieval. Recommender systems mostly run on revealed preference: what you watched, what you skipped, how long you stayed, what people with similar behaviour did next. Nobody asked you anything. Your taste is inferred from your actions, which is why the home feed can know you want car restoration videos before you would have thought to say so.
A prompt-built feed adds stated preference: you write the sentence, and the system is expected to honour it. The research literature on this is clear that neither signal is strictly better. Behavioural signals are cheap and continuous but carry every bias in the book — you watched it does not mean you wanted to. Stated preferences capture what a person would endorse on reflection, but people are imprecise about their own taste and sensitive to how the question is framed. The interesting designs blend the two.
The clearest published example of that blend is CTRL-Rec, a 2025 paper from Micah Carroll and co-authors, which trains models to predict whether a user would approve of an item given a natural-language request, then folds those predictions into a conventional recommender's existing signal weighting — with a weight that sets how much the stated request overrides inferred behaviour. In their user study with Letterboxd users, natural-language control improved people's sense of control and their satisfaction without reducing engagement. That is the shape to have in your head for Custom Feeds: not a search engine bolted onto Home, but the normal ranking stack with a sentence added as one more heavily weighted signal.
Which explains both of the feature's requirements. Watch history has to be on because your behaviour is still doing most of the work. And the prompt is a filter and a tilt, not an instruction set — the feed is still a recommendation, with all the ordinary ranking considerations behind it.
YouTube has not said how videos get chosen, and that matters
There is no published documentation on what a custom feed reads from your video. No confirmation that it uses your title, your description, your transcript, your topic classification, or some embedding of all of it. No creator-facing control, no opt-in, no setting in Studio. Anybody publishing a checklist of "custom feed ranking factors" right now has invented it.
What you can reason about safely is the shape of the problem. A system matching a sentence like "witty podcasts for my daily 30-minute commute" against a catalogue of billions of videos has to turn the sentence into something comparable with each video's representation, and whatever that representation is built from, it is built from text and signals YouTube already holds about the video. The practical consequence is unglamorous and familiar: a video whose subject, format, level and length are stated plainly somewhere in its metadata is easier to match than one whose subject is only legible from watching it. That was already true for search. It is now true for a slice of browse traffic too.
This is the same conclusion the answer-engine surfaces pushed us towards, and for the same reason: a retrieval layer that reads text cannot see your image. Two audiences now decide the fate of one upload — a text-only one that decides whether you get found, and a human one with eyes that decides whether you get watched. Custom Feeds add one more surface where both get a turn, in that order.
What a sentence can ask for that a tag never could
The genuinely new thing about prompt-built discovery is the kind of constraint a viewer can express. Chips and search queries are topics. Sentences carry qualifiers, and the examples YouTube and the press have used are full of them: a duration ("under 10 minutes", "my 30-minute commute"), a mood ("help me unwind", "witty"), a stage of life or skill, a recency, a language, a format.
Some of those qualifiers are checkable against facts YouTube holds about your video, and some are not. Length is checkable. Language is checkable. Format is usually inferable — a podcast episode, a tutorial, a vlog. Mood is barely checkable at all, which is exactly why mood prompts will produce the loosest feeds and the most "Something wrong?" taps. The useful move is to make the checkable things true and unambiguous rather than trying to write for the vibe:
- Say the format in the title or the series name. "Episode 14" tells a system nothing. "Beginner piano lesson 14" tells it the format and the level.
- Let length be a feature where length is the constraint. If a video genuinely is a 12-minute full-body workout, saying so costs you nothing and makes you matchable against every prompt with a time budget in it.
- Put the subject in the first line of the description, in the words a viewer would use. Not keywords — a sentence. The description's opening is the part every retrieval layer has always weighted most.
- Caption accurately. An accurate transcript is the cheapest large block of honest text about your video that exists, and the one most creators leave to automatic captions without checking.
- Keep series naming consistent. A feed assembled around a recurring interest is a feed where recognisable series names compound.
None of that is new advice, and that is the point. The honest version of "optimising for Custom Feeds" is ordinary metadata hygiene done once, properly, rather than a new discipline — the same work that has paid off in search for a decade, pointed at one more surface.
The thumbnail's job changes state, not importance
A custom feed is still a grid of thumbnails. Nothing about the feature removes the image, shrinks it, or puts text where a picture used to be. What changes is the mental state of the person looking at it, and that is a bigger deal than it sounds.
On the ordinary home feed, the viewer has no declared intent. They are idle. Your thumbnail's job is to create a reason to care out of nothing, which is why curiosity gaps, withheld information and implied stakes work there: the image has to generate desire, not satisfy it. That is the mechanism the whole psychology of clickbait rests on, and it is not going anywhere, because the main feed is not going anywhere.
Inside a feed the viewer built from a sentence, the state is reversed. They have already declared what they want. They are not asking "why should I care about this" — they are asking "is this the thing I asked for?" Against that question, mystery is friction. A thumbnail that withholds the subject to create intrigue forces the viewer to work out whether it even matches their prompt, while the video below it shows them plainly that it does. Fit wins, and fit is read off the image before anybody finishes the title.
Practically, that means a prompt-driven surface rewards the things that get called "boring" thumbnail design: the subject visible and identifiable, one clear claim, the format obvious at a glance, type you can read at feed size. It rewards specificity over escalation. A thumbnail for a guided-meditation video that reads "10 MIN · SLEEP" in a feed about unwinding after work is doing more work than a face in shock, because the viewer's question is literally about match.
Your competitive set shrinks, and that changes what differentiates you
The second-order effect is subtler and probably more important. On Home you compete with everything — a different niche, a trailer, a football highlight, somebody's cat. In a feed assembled around one sentence, you compete with a tight set of videos that are all plausibly the same thing.
Genre convention is what gets you into that set; it cannot get you picked inside it. If every thumbnail in a feed about beginner guitar uses the same close-up of hands on a fretboard with a yellow arrow, the convention has become invisible — it carries no information, because everything has it. What differentiates inside a narrow set is being recognisably someone: a palette, a type treatment, a crop, a framing that a viewer has seen before and can attribute without reading the channel name. That is the argument for thumbnail consistency made sharper, not softer, by prompt-built feeds.
It also quietly raises the value of a back catalogue that looks like a set. A feed that keeps refreshing against a standing interest is a repeat-exposure surface, and repeat exposure is where a consistent look pays compound interest instead of simple interest.
The living-room version
Custom Feeds were announced for TVs as well as web and mobile, and the television is where feed design is least forgiving. Thumbnails render large but are viewed from three metres away, the surrounding interface is sparse, and selection happens with a remote rather than a thumb — which means fewer items considered, and each one considered more deliberately.
Everything already true about TV packaging gets slightly more true here: fewer words, larger type, higher contrast, no detail that depends on a viewer leaning in. If you have never audited your own channel from a sofa — your own videos, on an actual television, from where you actually sit — a prompt-built shelf on a living-room screen is a good reason to finally do it.
Where this shows up in Analytics: nowhere, specifically
There is no custom-feed traffic source. YouTube's documentation defines Browse features as traffic from Home, the subscriptions feed, Watch Later, Trending and Explore, and other browsing surfaces — and a custom feed on Home is a browsing surface. Unless YouTube adds a breakout, which it has not announced, views from a prompt-built feed will arrive inside your existing Browse features line, indistinguishable from ordinary home-feed views.
That has one real analytical consequence worth planning for. Browse CTR will increasingly be an average of two populations whose click behaviour differs structurally: idle viewers with no intent, and viewers who declared an intent and are scanning for match. Blending them makes the number noisier and its movements harder to attribute. If your home-feed CTR drifts over the next two quarters without a corresponding change in your packaging, mix is now one of the candidate explanations, alongside all the usual ones.
The defence is the same as it has always been: track absolute impressions and absolute views by source, not just percentages, and never read a CTR number without the surface it came from. Traffic sources explained covers how to split that properly in the Reach tab, and what counts as a good CTR covers why a single-number benchmark was never meaningful in the first place.
The honest size of this
Two things keep the feature in proportion. The first is adoption. Research on feed-customisation tools is consistent that building a feed by hand is effortful and that only a small fraction of users on platforms offering custom feeds ever build one — a 2026 paper on continuously aligning social feeds makes exactly that observation while proposing a lower-effort alternative. YouTube's own 30-day prompt expiry is an implicit acknowledgement of the same pattern. Most viewers will not write a sentence. Most of your views will keep coming from a feed nobody configured.
The second is history. YouTube has shipped viewer-controlled discovery surfaces before. "New to you" launched as a chip at the top of Home in 2021, rolled out across mobile, desktop and TV, and is remembered by approximately nobody. That is not a prediction that Custom Feeds will go the same way — a prompt is a far stronger expression of intent than a chip, and the retrieval technology behind it did not exist in 2021 — but it is a reason to treat the first wave of coverage sceptically.
Set against that, the reason to care at all: recommendations are where the watch time is. The most-quoted figure on this is YouTube's own claim, made by Neal Mohan at CES in January 2018 when he was chief product officer and reported at the time by Tubefilter, that around 70% of watch time came from recommendations rather than search. It is a company claim with no published methodology and it is now nearly nine years old, so treat it as an order of magnitude rather than a statistic. The order of magnitude is the point: a change in how the recommendation surface works is a change to the biggest part of the funnel, even when the change starts small.
What to actually do in the next month
- Nothing structural. Do not reorganise your channel, your upload schedule or your niche around an experimental feature in one country in one language. The cost of being early here is real and the payoff is unmeasurable.
- Fix the metadata on your ten best-performing videos. Subject in the first line of the description in plain language, accurate captions, format and level stated where they are part of the value. An afternoon, once, and it pays into search and answer engines too.
- Name your series properly. If you make recurring content, the series name should describe the thing, not number it.
- Audit your last twenty thumbnails for attributability. Shuffle them with twenty from competitors at feed size and see whether a stranger could sort yours into one pile. If not, that is the gap a narrow feed will expose. The thumbnail preview tool renders them at real feed sizes for free.
- Check one video's packaging for "fit legibility". Cover the title. Can you tell what the video is, what format it is and roughly how long it feels, from the image alone? That is the question a prompt-built feed asks.
- Write down your current Browse features impressions and CTR. A baseline recorded now is the only way you will ever detect a mix shift later.
- Look at your channel on a television. Ten minutes, and it is the surface most creators have never actually seen their own work on.
What not to do
Do not write for the prompt. Stuffing mood words, time qualifiers and commute references into descriptions in the hope of catching sentences is the 2013 tag-stuffing playbook wearing new clothes, and it fails for the same structural reason: a retrieval layer that can parse a sentence can also tell that your description is not a sentence. Worse, a claim in metadata that the video does not honour buys you an impression and then a bounce, and the one thing every version of YouTube's ranking has agreed on for a decade is that it would rather not show a video people leave.
Do not buy a product that promises custom-feed placement either. There is no creator-side lever, no documented ranking factors and no reporting line. A tool cannot optimise what the platform has not exposed.
Frequently asked questions
Can creators see which of their views came from a custom feed?
No. There is no separate traffic source, and YouTube has not announced one. Those views sit inside Browse features alongside the rest of your home-feed traffic.
Do custom feed views count normally?
They are views on Home like any other, and coverage of the test describes them behaving as normal home-feed views for watch time, with videos watched there entering the viewer's ordinary watch history and feeding back into their main recommendations. YouTube has not published anything suggesting a different treatment.
Is it available outside the US?
Not yet. YouTube's Help Center describes the shipped version as experimental and available in English in the United States, and the Made on YouTube announcement scoped the full rollout to US viewers on web, mobile and TVs.
How many feeds can a viewer keep?
The earlier experiment allowed one active feed. Coverage of the multi-feed test puts the cap at eight. Either way, prompts expire after 30 days of inactivity, so the practical number any given viewer maintains will be smaller than the cap.
Does this make clickbait stop working?
No. It makes clickbait stop working in this one surface, where the viewer has already declared intent and is scanning for match. The main home feed, Suggested and the Shorts feed all still serve viewers with no stated intent, and that is still where most of your impressions come from.
Should small channels be more or less interested in this?
Slightly more, in principle. A surface that retrieves against an explicit description has less need of the behavioural history that advantages large channels, and narrow, clearly-described content is exactly what a specific sentence asks for. But "in principle" is doing a lot of work in that sentence, because nothing about the actual selection has been published.
Where this leaves the work
Custom Feeds are the first real crack in an assumption that has shaped creator advice for over a decade: that the viewer arrives silent and the packaging has to do all the talking. In a feed built from a sentence, the viewer talks first. That does not demote the thumbnail — it is still a grid of images, still the thing scanned before anything is read — but it changes what the image has to accomplish. Less manufacturing desire, more proving match.
The uncomfortable implication for most channels is that both jobs now have to be done by the same catalogue of images. The main feed still rewards a thumbnail that creates a reason to care. A prompt-built feed rewards one that answers a question the viewer already asked. The channels that handle that well will be the ones whose thumbnails were specific and recognisable to begin with, because specificity reads as intrigue to an idle viewer and as fit to a deliberate one. Vagueness reads as nothing to either.
Which makes the practical constraint the same one it always is: producing enough good, consistent, specific images that your channel is recognisable at feed size on any surface YouTube invents next. That is a production problem rather than a creative one, and it is the problem Thumblore exists to remove. If you want the rest of the September announcements in context, they are in the Made on YouTube 2026 breakdown; if you would rather spend the afternoon on something with a guaranteed return instead, the rules that actually move click-through rate have not changed at all this year.