What 78,181 scanned followers told us about Web3 community

We scanned all 78,181 followers of one Web3 project. 90.96 percent failed authenticity screening. The 5,806 real people left over told us more about that community than any dashboard ever had.

This post walks through the first CommunityOS production scan — the Mintlayer deployment — number by number. Not because the numbers flatter anyone (a 91 percent filter rate is a confronting result for any project), but because they are exactly the numbers most community reporting is built to avoid.

What did the scan actually measure?

The pipeline runs in two stages. First, Bot-Kill — a pre-scoring filter that removes accounts showing the signatures of farms, shells, and inactive husks before any scoring happens. Second, the scoring engine itself: every surviving account is classified across four archetypes — Champion, Amplifier, Builder, Early Adopter — using a deterministic model weighted 60 percent on linguistic features and 40 percent on normalized vanity metrics. Deterministic means the same inputs always produce the same outputs; there is no model temperature, no sampling, no "it scored differently on Tuesday." The full method is documented on the engine page.

The output is not a report. It is a ranked queue: of the 5,806 real accounts, 298 surfaced as worth an operator's attention that week, each with the reason it ranked and a suggested next action.

Why did 90.96 percent get filtered?

Because that is what the follower list of a Web3 project that has been through airdrop cycles actually looks like. Airdrop farming, engagement farms, and follow-for-follow rings inflate Web3 follower counts in a way that most industries never experience — the incentives attract automated accounts at industrial scale. The dynamic is a cousin of what security literature calls a Sybil attack: one operator, thousands of identities.

Two honest caveats. The 90.96 percent figure is one project's result, not a universal constant — audiences that never ran incentive campaigns filter lower, and established consumer brands materially lower still. And Bot-Kill is deliberately conservative in what it lets through: the cost of a farm account reaching an activation campaign is higher than the cost of a quiet real account being held back, so ties break toward filtering. The four signals it uses are the subject of their own teardown.

What does a community of 5,806 actually look like?

Smaller and far more useful than a community of 78,181. Once the noise is gone, structure appears. The real accounts split into the four archetypes, and the split is the operating insight:

• Champions — the voices. They speak about the project with depth and conviction, and people who trust them listen. There are never many. Every one of them matters.
• Amplifiers — the reach. Wide audiences, fast distribution. They carry a message into rooms the brand cannot enter.
• Builders — the substance. They ship write-ups, tools, and threads with original work in them.
• Early Adopters — the first wave. They show up before it is obvious and give the honest first signal.

The ranked queue that falls out of this is the entire point of the exercise. On scan week, the operator opened a list of 298 people ordered by why they mattered right now — a Champion who posted twice about the protocol that week, an Amplifier who engaged the pinned announcement, a Builder whose write-up referenced the docs. That is a to-do list, not a dashboard.

Which numbers should you actually report?

Here is the uncomfortable part. If 90.96 percent of a follower base is inauthentic, then every metric computed over the raw base — engagement rate, impressions-per-follower, growth rate — is computed over fiction. The denominator is broken, so every rate built on it is broken too.

A follower count is not a community metric. It is a mixture of community and noise, in unknown proportions — until you measure the proportions.

The numbers worth putting in front of stakeholders are the ones that survive verification: how many real people are in the audience, how many were activated, how many completed actions that passed Proof Review. That reporting stack is what agencies use CommunityOS for — a community ROI number a client's CFO can interrogate without it falling apart.

What changed after the scan?

Three things, and they generalize to any project that runs one. First, outreach stopped being random: the queue replaced "who should we talk to?" with a ranked answer. Second, reporting switched denominators, from 78,181 nominal followers to 5,806 verified members — every subsequent percentage became meaningful. Third, the project stopped paying attention costs on noise: no more drafting announcements calibrated to an audience that was 91 percent shells.

The full case study, with the campaign that followed, is at /case-studies/mintlayer.