How the score is calculated

What we measure

Six signals adding up to 100 points. The weight is the ceiling for that signal: engagement can never contribute more than 30 points, trust never more than 8.

The “Similar” block

Below a listing we show up to five entries it is connected to most often: mutual mentions in posts, forwards and links in the bio. This is not a quality rating and not our recommendation — it is what the data shows. We only list entries from the catalog, of the same type and in the same language.

How we measure the rest

Beyond the score, pages carry a few more numbers and blocks. Here is what each one means and where it comes from.

• Chats. We count members and how many of them are online — the public chat page reports both. Chats have no score: the score is built from views and posting regularity, and a chat has no posts.
• Bots and mini apps: “Monthly audience”. The figure Telegram shows inside the bot itself — how many people used it over the last thirty days. It does not compare with a channel's subscribers: a subscriber stays on the list, a bot user counts only when they come.
• “Similar in meaning” in search. When exact matches are few, we show entries whose content is close to the query. Closeness is computed from a numeric representation of the text — title, description and recent posts — and it is a hint, not a match.
• “Network”. Sometimes several entries turn out to be linked so closely that the data shows it: nearly identical content, or links in the bio shared by a handful of entries. We call that a network and list its members. It is an observation, not an accusation: a network often has an honest reason — one author running several channels openly.
• “Audience: country”. Appears where the readers' country differs from the majority in that section. It is computed from language and from words in posts, not from data about people: Telegram does not publish audience composition, and we will not fill that gap with guesses.
• “Ads: about N a week”. How many advertising posts we counted in a channel over recent weeks. The markers are the author's own label, a link to another channel in a recognisable shape, and the same offer repeating. This is our reading of the posts, not a statement by the owner.

Where the numbers come from

Everything we know about a channel comes from its public t.me page: subscribers, posts and view counts. We look once a day and keep one row per channel per day. We store post texts in full — search and analysis rely on them; lists show the first 300 characters, the whole text appears on the post page.

Limits worth stating up front

• The monthly post count comes from the whole history we have collected, not from a single pass: an active channel can run to two hundred and fifty. Before the day a channel entered the catalog we have no history for it.
• Reactions and forwards come from the Telegram protocol — no bot needs to be connected. The signal switches on from the seventh day of observation: over a couple of days the ratio is understated, because reactions accumulate along with views. Until then we say “not enough data” instead of writing a zero.
• History starts the day a channel enters the catalog. We know the channel's own age from its earliest post, but we can only chart momentum from our first measurement onward.

When data is thin

What we did not measure is not counted as zero — it drops out of the calculation, otherwise a channel would lose points for our own blind spot. The score is then worked out from what we did measure and brought back to the same hundred. On the channel page such a row is labelled “not enough data”, with the share the score actually rests on next to it.

If less than 60 points of weight are measured, or we have watched the channel for under a week, the score is flagged preliminary. It is still shown, with the flag: precision arrives with history.

Anti-fraud

Bought subscribers look nothing like organic growth: they arrive in rare sharp spikes against a flat background and often drain away over the following days. We track the share of such days inside a 30-day window, and separately penalise an influx followed by an exodus.

The second line is the gap between audience and views. Subscribers who never read push engagement well below the cohort median, so that signal drops on its own without a dedicated rule.

A confirmed fraud flag halves the total; violations confirmed by a moderator take off up to 40 points. We do not publish the exact trigger thresholds: the method is checkable, but a manual for getting around it does no reader any good.

The third layer is views that cannot exist. A median view count well above the subscriber count happens to an honest channel after one post travels far, but it does not hold for months: the audience is the same size, yet twice as many people appear to read it. We do not treat that engagement as outstanding — the signal drops to zero, and the channel page says why. The bar is set per size cohort: for a small channel, a single forwarded post honestly brings more views than it has subscribers.

Algorithm versions

Every stored score records the algorithm version that produced it. Recalculation does not rewrite the past, so the chart on a channel page stays honest across revisions.