OpenKol
About

We built the thing
that told us we were wrong

This started as an attempt to rank KOLs. The data refused, and the product is what was left once we stopped arguing with it.

Read the methodology
Why

A market that moves real money on a metric it can't check

Projects pay accounts to post. The rate is set by follower count, and the outcome is assessed by looking at the replies. Both of those are purchasable, and both are routinely purchased.

The gap
There is no accepted way to answer “did that post do anything”, which means there is no accepted way to price the next one. Everyone in the transaction knows this and nobody can fix it alone.
What we do
Measure what happened after a post: the volume it brought in, whether the engagement was real, and how both compare to everything else we have measured. Retrospective, consistent, and checkable.
What we don't
Predict. Rank people. Sell a better score. Each of those would be more commercially obvious than what we actually built, and each would be a claim we can't support.
What we got wrong

Three assumptions the data killed

Every one of these felt obviously true before it was measured. They are on the site because a measurement tool that only publishes its successes is asking you to take its failures on faith.

Good KOLs stay good

Splitting each account's history in half and correlating gives −0.024. One went from +22.8% to −10.7% between its own halves. There is no leaderboard on this site because the data will not support one.

A real audience means good calls

Audience quality against subsequent returns: 0.052. A real audience is worth knowing about and tells you nothing about the token. So the scores stay separate rather than being averaged into one flattering number.

Bot accounts look like bots

Scoring engagers on follower count, age and ratio was inverted on labelled data. What actually works is reading the replies. That took the classifier from 78% to 91%.

How we work

Rules we set before they were inconvenient

Each of these costs something. They were decided in advance for exactly that reason.

Withhold rather than estimate
Where a component cannot be computed we leave it out and name the gap. A withheld number is honest; an estimated one is a fabrication with a decimal point on it.
Publish the failures
The measurements that came out badly are on the methodology page, in their own section, not in a footnote. They are the reason to believe the rest.
Never sell a score
We sell publication of the real number, never a change to it. The moment placement is purchasable the whole thing is worthless — and everyone here will assume that is what is happening unless the policy is loud.
Nothing accusatory in public
Unvalidated signals stay private to whoever ran the check, are marked experimental, and always show their evidence. Labelling a named person on a detector with no confirmed cases behind it is not a risk worth any amount of attention.

Measurement, not predictions.

It is a smaller promise than the one this market is used to hearing. It is also the only one we can keep.

Get in touch

Check one post.Stop paying for reach.

Read the methodology