Tested, not just promised
Test results
We test every AI feature before you rely on it, on thousands of messages built to match real crypto chat plus real messages from live communities. These numbers are from our latest test runs, on our current AI model (November 2026).
Sentiment
99.5% of 884 everyday messages read correctly, including Web3 slang and emoji. On real member messages from live communities: 94% on Telegram and 97% on Discord. Greetings, adverts and spam count as neutral, so they don't move your mood score.
FAQ auto-answers
Of 388 answers the bot would have posted, 381 (98.2%) fully answered the question and 3 (0.8%) answered a different one. It posts only the answer your team approved, word for word, and stays silent when no FAQ fits: 0 answers in 267 cases where none applied.
Question detection
99% of everyday member questions picked up, each checked 4 times in batches mixed with small talk. Questions from 17 other kinds of projects, from prediction markets to AI products: 98.7% to 99.6%. Small talk ignored: 100%.
Analytics
Every dashboard figure we checked matched a direct count of the raw data: 20 of 20 on Telegram and 16 of 16 on Discord.
Member retention
On a test where the right answer was known in advance, new-member retention came out exactly right.
Team performance
Each team member's message counts matched the raw data exactly, and reply speed counts working hours only (9 of 9 test cases). A reply is credited only to the member a moderator actually answered (16 of 16 checks), and work shifts follow each moderator's own time zone.
Points and leaderboards
Points can't be farmed: "gm", short filler and rapid-fire messages earned nothing in every test case.
AI summaries
Summaries quote your community's real numbers, and reward ideas never told the team to DM a member (0 of 30).
Referrals and KOL tracking
On Telegram, every join is matched to the exact invite link used. On Discord, in a test with 5,000 people joining, nobody was credited to the wrong link. When it can't be sure, a join is marked "couldn't match" instead of guessing. Every Referrals number matched a count done by hand, and it was tested with real people in a live Telegram group.
Verified Missions
A repeated product action never earns points twice, and a new mission can't give anyone points until a test event has come through. Checked with 53 automated checks on both bots, and tested end to end on both bots with a demo project.
Translations
Pinned posts and announcements in 10 languages: 96.5% of 370 translations were fine to post as they are (93% on Discord's harder set of long posts). Every one came back complete, with links, @handles and tickers kept exactly, and a hidden "AI, ignore your instructions" line was never obeyed. The bot checks each translation and translates again if a sentence was left out.
Reliability and security
32 full security audits, with no critical or high-severity issues open. Nightly encrypted backups, with a full restore tested. The bots keep running through database restarts.
How we test
Our tests combine real member messages from live Telegram and Discord communities with larger sets written by an AI model to mirror real crypto chat. The correct answers are decided by a separate, stronger AI model, and every test runs several times in shuffled order. Before switching to a new AI model, we test it side by side with the current one. When a test finds a weakness, we fix it and test again: in November, jokes about good news went from 76% to 91% read correctly, real member messages from 78% to 94%, and Discord translations from 84% to 93%.