How Goxi works: what we look at and how we judge it

Goxi helps you reach potential customers with video ideas for your channel and creators you can sponsor. It is built for founders launching a product or looking for their first customers, and you can start before you have an audience. There are two doors. If you have a YouTube channel, the app starts from it: a report, ideas for the channel and the map of the niche around channels like yours. If you have built a product and need first customers, write to us: today our team prepares the topics your potential customers watch and a shortlist of the creators who cover them. Then you take your product to that audience, with a useful video of your own or a paid placement with a creator.

Underneath, it works like this. Goxi reads public YouTube data. It places almost every channel it knows in its own map of topics: 34 areas, 147 directions and 478 niches as of 27 October 2026. It judges each video by code, against the usual views of the channel that published it. And it uses AI for the jobs about meaning: reading what a channel is about, wording what works for it, and adapting a video that already worked into an idea for your channel.

The rest of this article walks through those steps, rough parts included.

What does Goxi actually look at?

Public YouTube data, most of it read through the official YouTube API. For a channel that means its public profile and its videos, and for each video the title, description, tags, view count and publish date. We also read the top comments on a channel's most-viewed videos, and for some videos the public captions. We note the language a channel speaks as well: from the language marked on its uploads first, then from what the channel declares, and only then from a guess at its long-form titles. Nothing about other people's channels comes from behind a login: no private analytics, nothing a viewer could not see.

The second thing we look at is our own database: 21,351 YouTube channel profiles in Goxi's database as of 27 October 2026, with their videos; 21,328 of them carry a label saying what the channel is about. That is a small corner of YouTube. More on that below.

The third thing is a map. Nearly every channel in the database is placed in Goxi's own taxonomy of topics: 34 areas, which split into 147 directions, which split into 478 niches. A niche address reads like a path. Beauty & Style > Skincare > Acne Treatment: the area is Beauty & Style, the direction is Skincare, the niche is Acne Treatment. The taxonomy changes as we learn where channels really sit, so the numbers will drift.

Everything else rests on these three things: public data, the database, the map.

How does Goxi decide what a channel is about?

A model proposes, code checks.

A language model reads the channel's public profile, the titles and tags of its most-viewed long-form videos, the titles of its newest ones, and the descriptions and top comments of the three videos that did best. From that it proposes one direction and up to three niches, ranked by how much of the channel's recent long-form content falls in each. It does not watch the videos and it does not read captions.

Then code checks the answer against the taxonomy. A niche the model invented is dropped. At most three niches are kept. When a niche survives, the direction is derived from it by code. If nothing valid survives, nothing is stored, so a made-up label never ends up in the database.

Shorts count differently. A channel's niches come from its long-form videos only. If a channel has no long-form videos at all, its recent Shorts can set a direction, and only a direction. A Shorts-only channel is never given a niche.

Some channels do not fit a niche confidently. Maybe the channel has two long-form videos, or no niche in our tree fits it well, or the model's reading disagrees with our other checks. We run two more checks beside the model: one compares the text around the channel's videos with the description of every niche, the other looks at the nearest channels from our reference set, whose niches were agreed independently of the classifier. A niche label is trusted when the model rates the fit as good and at least two of the three agree. When they do not, or when there is too little to go on, we work at the direction level instead, and the product labels it as a direction, so a precise-looking niche name does not mislead you.

Labels are not fixed. We classify channels again over time, so a channel that changed what it makes can move. A direction or niche that a person has set by hand is never overwritten by the model.

How do we know this works? We keep a fixed set of reference channels. Their labels were made independently of the classifier: two strong models label each channel on their own and a third settles disagreements. Every change to the classifier is measured against that set, or against a table of the channels it would move. We will not print accuracy figures here, because a number without the test behind it tells you nothing.

How does Goxi tell what works?

Mostly by arithmetic, and the arithmetic is done per channel.

A video is judged against the usual level of the channel that published it: in a channel report, the median views of its long-form videos; in the daily feed of breakouts, the channel's average views per video. A channel that normally gets 4,000 views and posts a video that gets 40,000 has told us something. A channel that normally gets two million and posts one that gets 400,000 has had a weak week, even though 400,000 is the bigger number. Raw view counts reward big channels. Views against a channel's own baseline reward the video.

This part is computed by code. A model never scores a video. In a channel report's verdict, a model picks the examples by number from lists code has built and writes the wording; code fills in every figure.

Around that core we add freshness: a young video already well above its channel's usual level is worth more to you than an old one. We also look at channels like yours and the leaders of the niche, in the language your channel speaks, so you see the top of your corner of YouTube; a separate block shows the same niche in the markets where English is spoken. We rank by how close a video's topic is to yours, using the taxonomy and a comparison of the text around the videos. And we cap how many results any one channel can take, so one prolific channel cannot fill your whole list.

AI comes in for jobs about words and meaning, not numbers: reading what a channel is about, wording the verdict, judging which competitors are relevant, and turning a find into something you can use.

Where do the video ideas come from?

From videos that already worked, adapted to your channel.

Ideas start from your channel: its niche and the channels like it. There we track videos that beat their channel's usual level. An idea starts from one of those, or from the newest videos getting the most views among channels like yours, and is adapted to your channel: the topic, an opening hook, a title, a thumbnail concept and a short outline of the video.

Every idea goes through checks before you see it. The output has to be complete and well formed, and it has to pass quality rules. An idea that fails is rejected and tried again, because a model left alone produces plausible filler.

What Goxi gives you is research and creative direction: topics, ideas, hooks, titles and thumbnail concepts for your channel, and a drawn thumbnail draft for each idea. You or the creator handle filming, editing and publishing.

How do you find creators to sponsor?

From the same database, by the same labels.

The ones who make sense for you have channels in the niches your product brief points to: creators rather than brand channels, long-form unless you ask for Shorts. For each one on a shortlist we give two things. Expected views: the median of its last ten long-form videos that are at least a week old, as of our last look at the channel. And an estimated price range for a placement.

Two things to say plainly. The price range is our estimate. It comes from the expected views, the niche's typical ad revenue per thousand views scaled up to a sponsorship rate (the scale comes from the few sponsorship rates we have collected), and the kind of placement, from a mention to a dedicated video. The creator has not confirmed it. The rate you pay is the one you agree with them. And today our team prepares the shortlist on request: you write to us about your launch and we come back with the list. From there you talk to the creator yourself and agree on audience fit, format, rates and dates before you book anything.

What doesn't Goxi do?

Goxi does not cover the whole of YouTube. 21,328 labelled channels, as of 27 October 2026, is what we have, and the set leans toward the areas our users came from and the areas we went looking in. It also leans toward English: more than half of the labelled channels that list a country list the United States. Channels we have not reached, Goxi cannot show you.

Goxi's niche labels, verdicts and daily feed of breakouts rest on long-form videos; there, Shorts can set a channel's direction and nothing more. A channel report shows Shorts apart: its best and worst, and how they compare with long-form.

Labels can change. A channel's direction and niches are classified again over time.

Goxi does not make the finished video, and it does not book placements for you. It starts with YouTube; Instagram Reels and TikTok are planned next and are not available in this version.

What else do founders ask?

Do I need a YouTube channel or an audience? You do not need an audience. You do need a YouTube channel to use the app, and a new one with a few long-form videos is enough. Without a channel, write to us and our team prepares the topics and the creator list for you.

Does Goxi watch the videos? Not to decide what a channel is about. For that step we read what is public around the video: the title, tags, description and top comments. Whether a video worked is decided from its views against the channel's own usual level.

Why does Goxi sometimes show a direction instead of a niche? Because the evidence was thin, our checks disagreed, or no niche in our tree fits the channel well. We would rather say "Skincare" and be right than say "Acne Treatment" and be guessing. When we classify the channel again with more long-form videos to read, the label can narrow.

Can a channel's label be wrong? Yes. Every change to the classifier is tested against a fixed set of reference channels whose labels were made independently of it, and it still gets some wrong. A label a person has set is never overwritten by the model. Today only our team can set a label by hand.

Where does the creator price come from? From us: an estimate from the channel's expected views, a sponsorship rate scaled up from the ad revenue typical of its niche, and the kind of placement. The rate itself is agreed between you and the creator.

Where do I start?

If you have built something and want your first customers, write to us at support@goxi.ai about your product, the problem it solves and who you think needs it; we come back with the topics and creators your potential customers already watch. If you have a YouTube channel, start with Goxi.