What every number means.

Each source subscore is half level and half momentum. Level asks how big the topic is right now. Momentum asks which way it’s moving, comparing the most recent 7 days against the 3 weeks before them. A perfectly flat topic scores exactly 50 on momentum, so a high total needs either real size or real acceleration.

The two are then combined by weight: Google Trends for 53% and YouTube for 47%, stretched to fill 100% because Reddit, X and TikTok are not connected yet. A source that fails on a search drops out the same way and the rest stretch to cover it, so one broken API costs you confidence rather than dragging the number toward zero.

The data comes from Google Trends through SerpApi, two requests per topic: one for the daily chart, one for the related queries. Days Google marks incomplete are dropped rather than charted as a dip.

The sample is the 50 most relevant videos published on the topic in the last 30 days, in YouTube’s own relevance order, and every view figure is a median rather than an average so one runaway hit cannot drag the picture up. A count on its own answers almost nothing: forty eight videos describes both a topic three channels shipped on last week and a topic whose best answer is fourteen months old from a channel that has since moved on. Age is what separates them.

The level half of the YouTube subscore averages three signals: how many videos, how fast the median one earns views, and how far it travels past its own subscriber base. All three use a log scale, so the gap between 500 and 5,000 views a day moves the score far more than the gap between 45,000 and 50,000. The momentum half compares uploads in the last 7 days against the 23 before them. Both halves are indexed rather than counted, so Upswing says which way a topic leans and roughly how hard, and refuses to turn that into a ratio of searches, which it is not.

Under the source cards is a second Google Trends reading over 12 months, folded into calendar months. Each bar is the average interest across that month, the brand coloured bar is the peak and the blue bar is now. The swing figure is the peak month divided by a typical one: below 1.6× the year is called flat and timing is not your problem.

It costs one extra Google Trends call, so it is kept 7 days rather than 6 hours and loads after the score. Under 10 months of history it draws nothing rather than guessing.

Nobody is asked who their competition is, because most people answer that question with the channels they admire rather than the channels taking their views. The list here is worked out from search: a channel that ranks for 2 or more of your own subjects is competing with you whether or not you had it in mind. Each row shows how many, so a result you disagree with can be removed rather than wondered about.

Every comparison is a video against what its own channel’s videos of roughly the same age normally do. Never your numbers against theirs directly, because a channel ten times your size beats you on every unadjusted figure, which is true and useless.

Age is the hard part. Dividing views by age gives an average rate across a video’s whole life, and a video has no rate: it takes most of its views in the first fortnight and trickles after. That buried every established channel and made three week old videos read as fifty times normal, so videos are compared within age bands instead.

The first pass is the one expensive thing in Upswing: each subject is a YouTube search at 100 quota units, so discovery costs roughly 800 of a daily 10,000. That is why it runs once and is stored, refreshes look only at what is new, and adding a channel by hand is nearly free.

What it cannot show you is why. Impressions, click through rate and retention belong to whoever owns a channel, and no tool can see them for anyone else. So a video here can be shown to have broken out, and never what made it.

Before Claude sees anything, the titles are measured: median length, how many use numbers, questions, brackets, a colon, “vs”, first person, or an all caps word. Both the full 50 and the top quarter are measured separately, so the difference between them is visible rather than imagined.

Your own catalogue is read the same way. Every upload is scored against what a video of its age normally does on your channel, using the age bands Competition uses, and your best and worst performers go to Claude with what each one actually did. Six real titles from your own channel, with real numbers attached, turn out to carry more than any summary of them.

An earlier version also tried to name the patterns behind that: which words, which lengths, whether model numbers helped. It was cut after being measured against four real channels, where it correctly found nothing on any of them. Word effects on real titles are small enough that on a catalogue of this size they cannot be told apart from luck, and a panel that says so on every channel forever is worth less than the six titles. What survives claims only that these videos did better, never that the wording is what made them.

A shared link is a snapshot and stays that way. It never re-fetches, so the numbers on it are the numbers that were there when you sent it, with the timestamp saying so.