Which number is right?

Each can be right for what it counts. Check the stage, the sample and the date.

Why is ARC's UCP share lower than some others?

ARC's sample and stages differ. Compare like with like. A profile count is not a checkout-link count.

Can I reproduce ARC's numbers?

Yes, from the CC-BY-4.0 dataset. Quote the version.

Why do agent-readiness checkers disagree?

Because they count different things in different samples. Some count a valid UCP profile. Others count a checkout link or a cart. ARC splits each protocol into declared, valid, usable and transacts, and publishes the sample, n and date, so a figure can be traced.

Short answer

Two checkers can both be right and still print different numbers. One may count a valid profile. Another may count a checkout link. The samples and the dates differ too.

Four stages

ARC saved checks, accessed 2026-10-13

UCP declared: 1,972 of 4,492 (43.9%)Measured/.well-known/ucp answered HTTP 200 with a bodySnapshot 2026-10-12.r2, 12 October 2026.

UCP valid: 1,922 of 4,492 (42.8%)MeasuredProfile passed ARC's validationSnapshot 2026-10-12.r2, 12 October 2026.

UCP usable: 1,762 of 4,492 (39.2%)MeasuredThe UCP catalog tool returned the tested productSnapshot 2026-10-12.r2, 12 October 2026.

UCP transacts: 1,563 of 4,492 (34.8%)MeasuredA checkout link came back. Not a completed purchase. This matrix is not the US-only weekly snapshot.Snapshot 2026-10-12.r2, 12 October 2026.

Samples differ

ARC's directory is a curated list plus stores people submit. Another project may crawl a larger set of domains and count valid profiles only. Compare the stage, the sample and the date. Do not compare a profile count with a checkout-link count.

Dates and method versions

The method changelog records when a definition changes. A number from an older method is not wrong. It is a different measurement. Quote the version.

How to compare fairly

Match the stage (declared, valid, usable or transacts), the market, and the date. The US weekly write-up and the Protocol Matrix are different samples on purpose. The open dataset is CC-BY-4.0, so the count can be recomputed.

Frequently asked questions

Why do agent-readiness checkers disagree?: Because they count different things in different samples. Some count a valid UCP profile. Others count a checkout link or a cart. ARC splits each protocol into declared, valid, usable and transacts, and publishes the sample, n and date, so a figure can be traced.
Which number is right?: Each can be right for what it counts. Check the stage, the sample and the date.
Why is ARC's UCP share lower than some others?: ARC's sample and stages differ. Compare like with like. A profile count is not a checkout-link count.
Can I reproduce ARC's numbers?: Yes, from the CC-BY-4.0 dataset. Quote the version.

On ARC

Cite this: ARC, arcreport.ai, accessed 2026-10-13, index snapshot 2026-10-12.r2, CC-BY-4.0. Protocol figures: Live saved checks, accessed 13 October 2026. Not a frozen snapshot. Index browser figures: Snapshot 2026-10-12.r2, 12 October 2026. Platform figures: Live saved checks, accessed 13 October 2026. Not a frozen snapshot.

Every figure is labelled Measured, Inferred or Not measured. Not measured is not zero. A missing result is not a mark against a store. ARC stops before payment.