BITCOIN ENTITY CLUSTERING

What this page is for

Maketo groups Bitcoin's addresses into owners, then counts and measures those owners. This page says how the grouping works, how well it currently works, and the things we refuse to claim with it.

If you came for the numbers rather than the method, they are on the owners page.

Why counting addresses tells you almost nothing

An address is free to create, and most wallets make a fresh one for every payment. So one person can sit behind ten addresses, or ten thousand.

Add addresses up and call the total holders, and you have counted labels rather than people. It always comes out too high, and there is no way to know by how much without doing the grouping.

The one rule that does most of the work

When a single Bitcoin payment spends coins from several addresses at once, whoever signed it held the keys to all of them. That is not a guess about behaviour. It is how a payment is built: you cannot spend from an address without its key. So every time we see it, those addresses go down as one owner.

We read every transaction since the first block in 2009, on our own Bitcoin node, and we keep the evidence rather than only the answer. That second part matters more than it sounds. It means changing our mind about how to judge a piece of evidence is a rebuild that takes minutes, instead of a re-read of the whole chain that takes days. A decision that costs days gets made on patience. A decision that costs minutes gets made on evidence.

The second rule, and how carefully we use it

When you pay someone you rarely have the exact amount, so your wallet sends the leftover back to you. Spot the leftover and you have linked the payer's own addresses too.

This one is a judgment call rather than a fact, so we made it hard to trigger. We accept a link only when two independent signals agree, and one signal that used to be allowed to decide on its own is now only allowed to narrow a decision that something else already made. It was demoted for a specific reason: when it guesses wrong it does not merely stay quiet, it points confidently at the wrong output.

That change cut the links this rule produces by roughly 95%, and we kept the cut. Fewer links we can stand behind beats more links we cannot, because a wrong link does not stay small. It welds two owners into one, and every number built on top of it inherits the mistake.

Two readers we built and deliberately left switched off

The first follows long chains of payments where a wallet passes its leftover forward again and again. An early version picked the wrong output as the leftover, which had the effect of linking people who paid to the people they were paying. It now measures how long those chains run and groups nothing at all.

The second reads wallet fingerprints, the small habits that give away which software built a transaction. It works, and it stays off, because millions of separate wallets share a fingerprint. Same software is not same owner.

The part of Bitcoin we can group, and the part we cannot

An address only enters the grouping if it ever revealed a link, by co-signing a payment or by being caught as a leftover. Most addresses never do. Well under half of every address the chain has ever seen is in the grouped set.

This is the number that quietly goes missing elsewhere, and losing it always flatters the result. So every figure we publish off this grouping carries the share of the chain it actually covers, and where a figure has a chain-wide reading as well as a grouped one, we show both.

The owners page does exactly that. It prints how much counting addresses overstates the crowd among the addresses we can group, and next to it the floor that holds if every ungrouped address were its own separate owner. The first number is much larger than the second. Publishing only the first would have been the easy, misleading version, and it is the reason that page waited.

How we grade ourselves

We wrote the passing marks down on 4 August 2026, before the first full measurement existed, and we chose them from what the chain has to look like if the method is working, not from what our results turned out to be.

The file holding those marks is fingerprinted inside the code, so moving one shows up in our history as its own deliberate change with a reason attached. It cannot be nudged quietly. Writing the marks down before there was anything to grade is the entire point: afterwards, every threshold is negotiable.

Five things get graded:

• How often the method gives up and answers cannot tell. A method that abstains freely is not being careful, it is being useless, and everything downstream inherits the emptiness.
• How much of the eligible chain the grouping rule was allowed to learn from.
• Whether the privacy detectors catch what they are built to catch, without flagging everything in sight. A detector that flags everything scores perfectly on catching things.
• Whether the method is genuinely telling transactions apart, rather than answering the same way every time. An answer that never varies satisfies any average.
• Whether the picture still holds when you weight by coins instead of by number of transactions. Something can look healthy per transaction while failing exactly on the large ones, which are the ones a balance is made of.

The mark we are failing right now, and why we are telling you

The coverage mark is set at 80%. Measured across the whole chain in August 2026, we are at about 67%. Weighted by coins rather than by number of transactions it is about 69%, close enough to the first figure to say the shortfall is not hiding in the large transactions.

One cause dominates. There is a privacy technique that makes one person's payment look like two people co-operating, and our guard against being fooled by it is deliberately set far too wide. It refuses roughly a third of everything it examines, including a great many ordinary payments from years before that technique was invented.

Narrowing that guard is real work inside the part of the code we are most careful about changing, so until it is done the mark stays failed and stays published. Loosening the mark to fit the result would make the grade meaningless, and stopping that is exactly why the marks were written down early.

The one enormous group

Grouping the whole chain produces one group far larger than any other, holding roughly one in seven of all the addresses we can group. We checked whether that is real or a mistake, and the honest answer is mostly real.

Businesses that hold coins on other people's behalf genuinely connect to each other, through shared custody, through settling between themselves, and through payment processors. Their addresses genuinely do co-sign. Every published study that groups Bitcoin this way finds the same giant. It also grew smoothly rather than appearing in one step, which is what a real network looks like and not what a single bad link looks like.

The consequence is a rule we follow rather than a caveat we mention. We never hand one big mixed group to one named business. Where a figure would need that, we publish it only for the narrow, individually checked part, or we do not publish it.

Where names come from, and why no page shows one

Names never come from us guessing. They come from information that was already public, mostly organisations publishing their own addresses.

Today those names are used to test the grouping rather than to label anything. If two separately named businesses land in the same group, one of us is wrong, and that is a finding worth having. No page on this site currently attaches a company name to a group of addresses.

What we refuse to claim

Five refusals. They are absolute, not defaults we would revisit under pressure.

• We do not try to work out who a private individual is. We use names for organisations, where a public source already named them. An owner with no public name stays an anonymous number, permanently.
• We never treat a theft as evidence of who controls an address.
• We never treat a sanctions listing as evidence either. It is a legal fact about an address, not a statement about who holds what.
• We never treat an address that took part in a deliberate coin-mixing payment as belonging to one owner. By design nobody there controls the whole thing, so grouping it would invent a link that does not exist.
• We never treat the name a miner writes into a block as proof of which addresses that miner controls. Anyone can write anything there. Which blocks carried which name is a fact, and we publish it on the mining page. Who holds which coins is not, and we do not.

What never leaves our own machine

The grouping runs on our own hardware, and only summary numbers ever leave it. No address, no name, no wallet fingerprint and no piece of the underlying evidence is sent onward, stored in the site's database, or served to your browser.

A guard inspects every outgoing bundle for anything shaped like an address and refuses the whole bundle if it finds one, rather than removing it and carrying on. That guard has already stopped a bundle, and the thing it caught was harmless, which is roughly the behaviour you want from it.

How often it updates

Once a day. Each day's new blocks extend the same run rather than starting a fresh one, and the owner grouping itself is rebuilt from the stored evidence every night, so the published shape is never much older than the chain.

Blocks are only read once they are a hundred deep, so a short reshuffle at the tip of the chain cannot change anything we have already published.

Version and changelog

Current method: entity-v1.0. Meaningful changes to how the grouping works, and any movement in the marks above, get noted right here.

In short

This page explains how Maketo works out how many owners sit behind Bitcoin's addresses. It is a measurement with known limits, not a claim to know who anyone is, and nothing here is advice to buy or sell anything. Where the grouping cannot see, we say so rather than filling the gap with a guess.