MachineRead terminal-window signal markA stylized terminal window with a green signal scan-line and two data marks, indicating machine-readable signal inspection.MachineReadAI & search readiness audit

Blog

Why strict agent readiness is a separate lens

Understand why strict agent readiness has its own probe count and why changing audit scope can change the percentage without changing the site.

author
published
reading
4 min

A site can return readable HTML and still publish very little specifically for agent discovery. Conversely, an API product can expose several protocol documents while leaving gaps in its ordinary page structure. One percentage would obscure the difference.

MachineRead therefore displays strict agent readiness separately from Essentials. The strict score counts selected public discovery signals and carries its own denominator. Read it with the included, missing, and not-checked labels, rather than treating it as a general certification (MachineRead, "Methodology").

Essentials and strict readiness answer different questions

MachineRead's public methodology keeps three score concepts apart: the 100-point full rubric score, the 60-point Essentials Evidence Score, and the Strict Agent Readiness Score. The Essentials score uses a fixed 14-group evidence denominator for the public free-check surface. The strict readiness score uses an 8-probe default and can expand to a 21-probe full scope when optional protocol, account/auth, and commerce families are included.

Essentials asks whether the site exposes a fixed set of public evidence groups that help a machine read the site. Strict readiness asks a narrower question about explicit agent-native discovery and protocol signals selected for this audit scope.

Mixing the two makes the report less useful. A fixed Essentials denominator is good for stable comparison across the included rows. A strict readiness denominator is good for showing which agent-facing surfaces were actually in scope. The two can sit next to each other without pretending to be the same metric.

The denominator changes because the scope changes

The strict score is a percentage, but the percentage only makes sense with its denominator attached. MachineRead builds the result from probe labels included in the resolved scope. The default scope has eight labels. Protocols add six labels. Account/auth adds three. Commerce adds four. With all optional families selected, the maximum is 21.

That means a missing label and an excluded label are not the same thing. A missing label was in the resolved scope and was not observed under the audit conditions. An excluded label was outside the selected scope and appears in not_checked rather than being counted as a failure.

A practical example makes the difference clear. A default report can inspect eight strict labels and leave A2A Agent Card or OAuth metadata out of scope. An all-scope report can inspect those optional families as part of a 21-label denominator. The second report did not make the first report worse. It asked a broader strict-readiness question.

The API preserves the caveat instead of flattening it

The full response includes the strict percentage, earned and maximum counts, passed and missing labels, not-checked labels, and a caveat. A client displaying only the percentage throws away information needed to interpret it.

For example, five observed signals from a scope of eight and five from a scope of twenty-one describe different coverage. The count has not changed, but the set of questions has. Preserve the scope and denominator when storing results so later comparisons remain understandable.

Strict-readiness peer comparisons use that separate probe inventory. Do not compare their percentages directly with the 60-point Essentials evidence score (MachineRead, "Methodology").

How to read the strict score

Check the included families before comparing percentages. If a second run enables an additional protocol family, its denominator can change even when the site has not. Do not describe that change as a regression until you have compared the underlying labels.

A missing signal also needs an applicability decision. An ordinary brochure site may have no reason to publish an A2A card. An API service that intends clients to discover one has a more specific reason to investigate the absence. Choose scope according to the intended service, rather than adding irrelevant documents to improve a number.

Use Essentials to inspect its included page and discovery checks, and strict readiness to inspect the selected agent-facing signals. Neither score tests whether an agent can successfully complete a business task.

Works Cited

MachineRead. "Methodology." MachineRead, https://www.machineread.ai/docs/methodology. Accessed 10 Sept. 2026.

See also