Build1 publisher2 min readPublished
Telegram's median post reach drops from 31.4% to 3.3% as channels get bigger
An open dataset covers 3,205,346 public channels and scores each one as average views per post over subscribers. Its authors publish a per-band median because forwarded and promoted views push some ratios above 1.0.
The Engineer · Build desk
What happened
- A catalogue of 3,205,346 public Telegram channels was measured for post reach and the result published as open data.
- Each channel's score is average views per post divided by subscribers, and each size band's headline figure is the median of those ratios across the channels in it.
- The median post in the 1,000 to 5,000 subscriber band is seen by 31.4% of subscribers, while above 500,000 subscribers the median is 3.3%.
- The two largest bands are computed exactly over 8,380 and 1,211 channels, while the crowded small bands' percentile ladders come from a TABLESAMPLE and are flagged exact: false.
- Category counts were kept off the public page because only 719 of 15,240 sampled channels carried a label at all.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Views include forwards, cross-posts and paid promotion, so the ratio measures delivery and not the share of subscribers who saw a post. A channel scored above 1.0 is being read by people who are not on its list.
- decision A buyer pricing a sponsored post gets a peer figure to check the seller against. On the note's own example, a 200,000-subscriber channel at 12% is doing twice as well as half its size peers, and one at 2% is the one to question first.
- contradiction The size projection and the exact tables disagree on how many channels clear 100,000 subscribers. The note tells readers to take the exact count, so anyone citing the projected figure is citing a number its own authors discount.
- capability Because the API ships each band's 101-point percentile ladder with sample sizes and rebuild timestamps, a quoted percentile can be checked against the population it came from instead of taken on trust.
A ratio above 1.0 is common on Telegram, and the note is explicit that it is not proof of a real audience [11]. Forwards, cross-posts and paid promotion all deliver views from people who are not on the subscriber list; channels in that state carry the verdict `amplified` [11]. A few channels several times above 1.0 move a band mean a long way. The median does not move [12]. Subscriber counts padded with bought or long-dead accounts push the ratio down, and a band mean would be dragged by those as well [13]. The authors write that the two distortions "do not cancel in any way you can rely on" [14].
Rows that get no score are labelled instead. Channels under 1,000 subscribers are excluded because a single forwarded post can put the ratio anywhere, and the API returns them as `below_floor` [8]. Groups return `not_a_channel`, since Telegram publishes no per-post views for a group [9]. A channel the crawler has never seen a view count for returns `no_view_sample`, and none of the three is folded into the statistics [10].
The score does not touch the crawler's own engagement field, which is still being backfilled; reading a half-filled column would make the score exist for some channels and not others with no visible reason [6]. Dividing the two stored inputs, the authors write, "is duller and always answers" [7].
The gap between the two published bands is a factor of 9.5 [24]. Whether it transfers to a channel you care about depends on the numerator. Reach here counts views, which Telegram prints on every public channel post [26]. So a channel running paid promotion is scored partly on delivery from outside its list [11], and a channel with dead subscribers is scored against a count that still includes them [13]. A single channel's percentage fits either case.
Sampled bands are sampled because they are big: the 1,000 to 10,000 range alone projects to roughly 338,000 channels [18].
The projection is the part of the release I would not quote. A separate sample of 14,431 channels with known subscriber counts projects roughly 7,900 channels above 100,000 subscribers [19]. The exact reach tables hold 9,591 [20], about 21% more [25]. The note says a miss of about 20% is what sampling error looks like in a band that is about a quarter of one percent of the network, and that the exact count is the one to trust [21]. The country facet came back empty on the same sampling pass [23].
What to watch
- Whether the crawler's engagement field, once backfilled, changes the published ratios or only their provenance.
- Whether category coverage rises above 719 of 15,240 sampled channels enough to put the facet on the public page.
- Whether the size projection is rebuilt on a larger sample and closes the gap with the exact reach tables.