How the ranking works

Method v1.1, 2026-10-21.

What this is

AI agents have started to pay for things on their own. An agent that needs a web search, a piece of data or an answer from another AI can pay for it on the spot, often a fraction of a cent at a time. Many of these payments use an open standard called x402 and a digital dollar called USDC, on a public blockchain called Base. Anyone can look up these payments.

We add them up every week and publish a ranking of the tools AI agents actually pay for. It works like a best-seller list, but for the services agents buy.

Why we don’t just add up everything

The raw numbers are misleading. Some payments are not real purchases:

people sending money to themselves;
tiny test “pings”;
one account pouring money into a service to make it look popular;
schemes where thousands of accounts pay to collect rewards.

In the month to 18 October 2026, about three of every four dollars went to sellers with fewer than ten customers.

So before we rank anything, we filter. The rest of this page explains how, step by step.

Step 1: Find the payments

We collect every x402 payment on Base, starting from 2 July 2026. We count only payments in USDC and EURC (digital dollars and euros), handled by the companies that process x402 payments. Euro payments are converted to dollars at the European Central Bank’s rate for that day, and the report also shows each coin on its own. Their own fees and refunds are not counted as spending. When a payment passes through a middleman account on its way to the seller, we count it once, for the account where the money ends up.

Step 2: Work out who was paid

A payment goes to an account number, not to a company name. We match account numbers to companies using public catalogs of paid services, and by hand when we have proof. The best proof: we ask the service for its price, and its reply names the account to pay.

If we can’t tell who an account belongs to, we never rank it by name. Its money shows up on a separate “Unlisted” line.

Step 3: Remove what clearly isn’t a purchase

Four rules run automatically, the same way for everyone:

Practice money. Payments made on test networks, or in anything other than USDC and EURC.
Paying yourself. The buyer and the seller are the same, or the buyer is one of the seller’s own accounts.
Pings. Payments under a thousandth of a cent. The cheapest price any seller lists is a thousandth of a cent, so nothing real is lost. Without this rule, ten pings would be enough to fake ten customers.
Fewer than ten customers. A seller needs at least ten different paying customers in the last 30 days to be ranked. Below that, we list it separately as “below threshold”. It joins the ranking in the first week it reaches ten.

Step 4: Show what might matter, instead of hiding it

Some patterns are worth knowing but are not proof of anything. We keep those sellers in the ranking and add a label:

“1 buyer: 97%”: one customer pays almost all the money. This could be a new service with one big early customer, so we don’t remove it. You can see how much it depends on that one customer.
“checked”: we looked into this seller closely and found no sign of fake buying. It is not an endorsement.
“under review”: we are looking into it and haven’t decided yet.

Step 5: Take a closer look at anything odd

Some sellers get a closer look:

the typical payment is very large, over $100;
most customers are brand-new accounts that never buy anything else;
a lot of money goes to an account nobody can name;
someone reports a problem.

What we check:

Is there a real website and product?
Does the service itself confirm the account is its own?
What does it sell, and do the payments match its prices?
Where did the customers’ money come from?
Does the seller send the money back?

When we remove a seller. Suspicion is never enough. We remove a seller only when we find evidence of one of these:

the money goes back: the seller sends the money back to its “customers”;
the seller funds its own buyers: the “customers” got their money from the seller;
many accounts, one customer: the “customers” were all funded by one source, and counted as one they are fewer than ten;
not a purchase: by the seller’s own description, the payment buys a place in a points, rewards or leaderboard scheme, not a service.

New accounts alone don’t count as evidence. Some apps open a fresh account for every user, or even every payment, to protect privacy.

Who checks. An AI assistant does the checks and writes down everything it found. The project’s owner reads and approves every decision before anything is published. Every removal is printed in the weekly report, with its reason and the dollars it removed. A seller who disagrees can ask for a re-check.

Step 6: Rank

Every Monday we rank sellers by the money they received from real customers over the last 30 days, up to the end of Sunday. Each category has its own ranking, such as web search, data or AI models.

What the numbers mean, and what they don’t

“Spend” is the dollars agents paid a seller. It is not the seller’s revenue or profit.
Not included: card payments, other blockchains, and free tools, since a free tool leaves no payment to count.
Example, the month to 18 October 2026:
- $734,944 in payments passed the automatic checks on each single payment.
- $529,134 of it went to sellers with fewer than ten customers.
- $23,073 went to two sellers set aside after a review. One had funded nearly all its own “customers”; the other’s thousands of “customers” were all funded seconds before paying by the same three accounts.
- $182,737 remained. 334 named sellers made the ranking, and $47,202 went to accounts nobody could name.

Our promises

The same rules apply to every seller.
Every dollar we remove has a written reason.
No one can pay for a place in the ranking or to change a rank. If anything paid ever appears on the site, it will be labelled as paid and kept outside the ranking.
When the rules change, we publish a new version and keep the old one.
Every number can be rebuilt from public data with our code.