Methodology

Which communities are included

Public Telegram channels and groups linked as official by crypto projects on CoinGecko and on DEX listing sites, plus communities discovered through forwards and links from those. Communities with fewer than 50 members are tracked but not published. Private and invite-only chats are never read.

Members and growth

Member counts are recorded at least daily from Telegram's public channel page, and hourly for active communities read directly. Growth figures compare the current count with the count recorded 1, 7, or 30 days earlier in our own history. We do not use third-party historical data, so history begins the day a community was first tracked, which is shown on every page.

Activity and views

Posts per day is measured over the most recent 30 posts. For broadcast channels, view counts are read from public posts and re-read 1 hour, 24 hours, and 7 days after posting, since views mature over time. Views per subscriber divides average views by the member count. Groups do not have view counts on Telegram, so groups are measured by the share of members online and by message volume.

Member quality

An estimate of what share of a community's members look like genuine accounts, with a likely range rather than a single number. It is built from the accounts we see while reading a community: everyone who posts, everyone Telegram announces as joining, and, where the admins leave the member list open, a sample of the list itself. Each account contributes a small set of public properties: roughly when it was registered (Telegram issues account numbers in order, so the number dates the account to within a few weeks), whether Telegram has since deleted or flagged it, whether it has a photo or username, the shape of its name, and when it was last online. Those properties are weighted, and the weights are fitted from what actually happens to accounts later: accounts Telegram deletes or flags are the examples of fake, accounts still posting a month later are the examples of real. Nothing is labelled by hand and no judgement is made about any person.

On top of the accounts themselves we look at the shape of the community: whether accounts were created in one narrow window rather than spread across years, how old accounts were when they joined, bursts of joins that do not move the message count, how many of a burst are still present a month later, how many members are online and posting compared with communities of the same size, and, for broadcast channels, whether views arrive quickly like real readers or in steps days after publication like purchased ones. Accounts that appear across many unrelated communities are counted separately, because that is the signature of a member-selling service rather than a shared interest.

Every figure is compared against communities of similar size, because what is normal for a 400,000-member exchange group is not normal for a 2,000-member project chat. Where a community hides its member list, which most large ones do, the estimate covers the members who post rather than the whole membership, and the page says so: that group is the healthiest part of any community, so the true figure is likely lower, not higher. Only aggregates are stored and shown. Members are identified by a salted one-way hash, never by name, username or account number, and no member is ever listed or scored individually.

Trust score

A 0 to 100 estimate of how genuine a community's size and engagement are. Every community starts at 60 and each signal moves it up or down by a stated weight. Signals are scored against communities of similar size in our own index, so "views per subscriber better than 80% of channels its size" means exactly that. For channels: views per subscriber, how much views vary across posts (bought views are uniform), whether views keep growing after posting, forwards and reactions per view, growth smoothness, who forwards the channel, and age. For groups: how many members actually post, share online, spam score, join burstiness from join events, growth smoothness, and age. How much data exists sets a confidence level and caps the score: low confidence cannot exceed 75, medium 88, and nothing scores 100. Labels: 75 and above looks genuine, 50 to 74 mixed signals, under 50 likely inflated. The score is a statistical estimate, not a finding about any account.

Written pages

Released profiles carry an overview, takeaways, cautions and FAQs written by a language model from the page's own metrics and a sample of recent public messages. The model may only use figures we pass it, never quotes or names a member, and is refreshed weekly. The auto-generated paragraph under the cards states the source data and date.

Controversy score

For groups with at least 100 member messages in the last 7 days: the ten busiest reply threads are reviewed by a language model that marks which are genuine arguments (disagreement, accusations, complaints at the team, members fighting) and how heated each is. The score combines the share of threads that are arguments, their heat, and the complaint rate. Groups with no confirmed argument score 15 or below. Confirmed threads appear on the page with senders anonymised and handles, links and numbers removed.

Shared audience

Posters are stored only as salted hashes, so we can tell that the same person posts in two groups without knowing who they are. Groups sharing at least five posters in 14 days are shown on each other's pages. When more than half of a small group's posters also post in another group, the page says so: that pattern is typical of paid raid crews and shill networks.

Model-scored sentiment

For released groups, a language model classifies a weekly sample of member messages as positive, neutral or negative and names the mood. The score is the share of positive messages counting neutral as half. Where no model score exists yet, a word-list score is shown and labelled as such.

Community Ratings

How the people discussing a coin feel about it, across Telegram, X and Reddit, over the last thirty days. Each person counts once per coin per day, by what they mostly said that day, and a message copied word for word counts once. On Telegram only open community chats count: a project's own group is reported on its own because moderators remove dissent, and call, airdrop, listing and spam channels are left out and shown as the promotion share. No single community can supply more than a fifth of a coin's Telegram votes.

Approval is the share of opinions that are positive; complaints count as negative and are also shown as their own rate, and neutral talk is reported separately rather than mixed in. Each platform is measured on the log-odds scale and corrected for its usual tone, the platforms are combined by random-effects meta-analysis so none can dominate, and the result is pulled toward what is typical for coins with as much discussion, with the strength of that pull fitted from every coin each day (empirical Bayes). A coin is rated only when at least two platforms each carry five or more opinions, and never on X alone, where small coins' talk is dominated by coordinated campaigns. The page shows approval with a likely range, calls a coin warmer or cooler than typical only when its whole range sits on one side, and ranks by the low end of the range so thin evidence cannot top the list.

Every message on all three platforms is read by one multilingual model trained for this task, which labels the author's stance toward the coin as positive, negative, a complaint, a question, promotion or neutral; promotion is kept out of approval. On messages it had never seen it agreed with a much larger model about three times in four, and was weakest on Chinese, so treat close calls as close. Ratings describe how a coin is being talked about, not where its price is going. Methods: Evan Miller on ranking by rating, Robinson on empirical Bayes, DerSimonian and Laird on meta-analysis, Spiegelhalter on funnel plots.

Rising Communities

Rising ranks communities by growth that is unusual and confirmed by activity, not by raw percentage growth, which rewards tiny groups and bought members. Member growth is measured on the log scale between the snapshots nearest each end of the window, and compared with communities of the same size measured at the same moment and with the community's own usual pace, pooled so a new community is judged against its peers and an established one increasingly against itself. The same is done for activity: people posting in groups, views per post in channels.

A community's score is its member growth, capped by its activity growth: growth counts only as far as active members or views grew with it. A sudden jump that makes up most of the window's growth, many times the community's usual hourly movement, with no tracked channel forwarding it just before, is taken out of the growth. A community that loses much of a recent gain within a week is marked down. Communities must have gained members after these adjustments, show activity, and pass the trust and spam checks; very new communities wait until they have enough history. Methods: velocity against a baseline as in Twitter's trend detection, size-dependent growth spread from Stanley and colleagues, and research on bought likes and fake trends.

Language

Each group's messages are classified by script and common words. Sentiment is scored only when at least half the messages are English; other languages show the language and no score rather than a wrong one.

Spam score

For groups, 0 to 100 from the share of messages that repeat earlier messages or come from bot accounts, how concentrated posting is among a few accounts, and messages per poster. Communities with a spam score over 60 are excluded from activity rankings.

Lookalikes

Handles that resemble an official channel, by shared core name, small edit distance, or similar title, found through Telegram search and through channels that forward official posts. Listed so readers can verify before joining. A listing is not an accusation; some lookalikes are legitimate fan or regional channels.

Risk flags

Pattern rules applied to public posts: referral schemes, wallet-connect prompts, DM solicitations, urgency language, listing claims, guaranteed-return claims, airdrop claims, and seed-phrase requests. Flags describe post content, not intent.

Corrections and removal

Channel owners can request corrections or removal on the removal page.