Browse

Browse the register

4,208,309 entries we hold a subscriber measurement for, arranged three ways. Each of the three is a property we have measured — not a category anyone assigned. Topic, further down, is the exception: a language model’s classification, labelled as one wherever it appears.

A register that can only be searched is a register you have to already know the answer to. These are the three axes the data supports: how big a channel is, what language it publishes in, and how recently it last published. Every list is ordered by the latest measured subscriber count, largest first, and so is every topic page. A separate surface asks a different question — not how big an entry is but which way it is going: risers and fallers.

Every count on this page is a measurement with a time on it. Each one is a pass over the whole register — the size bands alone scan two tables of 3.8 million rows — so they are taken on an hourly cycle and read from the result, not recomputed while you wait. The figures below were measured at 9 Nov 2026, 02:10 UTC. The listings underneath them are read live, at the moment you open the page.

By subscriber size

Fixed half-decade bands — each spans a factor of 3.16 — and they are not this page’s invention: they are the same bands the published engagement baselines are computed over, so a band page and the baseline for that band always describe the same population. A channel is placed by the most recent subscriber reading we hold for it, and a superseded reading never places it anywhere.

Entries is what the register holds in that band. With a baseline is how many of them have enough measured posts to carry an engagement figure, which is always a smaller number and is the population the median describes. The under-1,000 band has no baseline at all: the baseline dataset’s eligibility gate starts at 1,000 subscribers, so nothing is quoted for it rather than a neighbouring band’s figure being borrowed.

By language

Derived, not declared. Telegram publishes no language for a channel and neither does the channel: channel.lang is null for every entry in the register. What is measured is the language of each individual post, and a channel is listed under a language only when that language is the most common one across its own posts of the last 30 days and accounts for at least 60% of them. A channel whose month is a 40/35/25 mixture is not listed under any language, because it is not a channel in any one of them. This is exactly the rule the engagement baselines use to build their language cohorts, so the two never split the same channel two ways.

Posts measured is how many posts in the window carried that language across all of its channels — the evidence the label rests on, not a count of everything those channels published. Baseline cohorts is how many subscriber bands have a published engagement baseline for that language; a band needs at least 300 channels in the cell before one is computed, so most languages have none. Languages under 200 entries are listed here with their counts but are not given a page of their own: at that size a page would be a handful of rows under a heading, which is the thin-directory pattern this register exists not to publish.

By how recently it published

Measured against our own last visit, not against the clock. The crawler records when it last read each channel’s posts; the age below is the gap between that visit and the newest post it found. Stating it any other way would claim knowledge of what a channel did after we stopped looking.

The last row is not a bucket. Those entries are ones whose posts the crawler has not read yet, and they are counted apart from the dormant ones on purpose: “we have not looked” and “it has published nothing” are different statements, and only the second is ours to make. 3,894,264 entries were last read more than two days ago, so their bucket may be a boundary out of date.

By topic

Inferred, not measured — read this before the table. Every other dimension on this page is something a second person could recompute by hand from public data and get the identical answer. A topic is different: an on-box language model reads a channel’s own recent posts and assigns the closest of 31 fixed categories, with a confidence score attached. That is a classification, not a reading, and this register treats the two differently everywhere — a topic is never shown inside a table of measured figures without saying so. Each topic below links to a list of its channels, which carries the same label and prints the model’s confidence on every row; how the classification works.

Coverage, as read. The classifier has read 3,273,155 of the 4,208,309 entries the register holds a subscriber measurement for — 77.8% of them. Its most recent classification was written on 9 November 2026. An entry counted in none of the rows below has not been read by the classifier; that is not the same as the classifier finding nothing to say about it. 6 model/prompt-version combinations have contributed classifications so far.

Other / unclassifiable is listed last and has no page. It is where the classifier puts an entry it could not place under any of the other categories, so it is the absence of a topic rather than a topic, and a list of it would group entries by what we could not tell about them. Every other topic links to its channels, largest first by measured subscribers. The Adult and Software piracy lists can be opened but are marked so that search engines are not offered them. The topic index has the classification’s coverage and provenance on its own page. A channel’s own classification, with the model and date that produced it, is shown on its own page.

What is deliberately not here

There is no browse by posting rate. The register does not hold every post a channel has published — for a typical entry it holds 16, one page of the public preview. “Posts in the last month” counted from that is a floor set by how deep our crawl reached, not a count of what the channel published, and banding channels by it would publish a property of the crawler as a property of the channel. Recency survives the same objection where a rate does not: the crawler always reads the newest page first, so the newest post we hold is the channel’s real newest post as at that visit, whatever sits unread behind it.

The register’s post coverage was last refreshed between 5 August 2026 and 9 November 2026, across 4,113,870 entries. How we measure.