NASA’s scan and the problem with mandatory-sounding AI checklists

nasa.gov · observed 2026-10-02T18:25:23.306Z

The finding in context

The saved nasa.gov assessment returned 91/100 technical readiness, high observation confidence and 58% evidence coverage. Its recorded llms-txt-not-found finding was informational. The report described llms.txt as an emerging proposal, not a universal requirement.

It would be easy to turn that line into an alarming headline. It would also be unsupported. This observation does not show that NASA is invisible to AI systems, that its public information cannot be retrieved or that it needs a transactional agent interface. A missing optional file is a narrow observation.

Start with the task, not the checklist

A public-information site serves different tasks from a paid API or a storefront. Readers may need accessible documents, reliable canonical links and clear provenance. They may not need machine-readable commercial quotes, payment challenges or a protocol for consequential actions.

Before recommending an additional surface, name the actual task it would improve and the audience expected to use it. Check the documentation of the relevant system rather than assume every emerging proposal is a requirement. Google’s guidance for its AI search features points website operators back to normal Search eligibility and sound technical foundations, not a special AI-only markup requirement.

Keep unknown, optional and not applicable distinct

An unknown check means the collector did not establish the fact. Optional means the signal is not universally required. Not applicable depends on the site’s intended capabilities. Those states should not be flattened into red warnings, and a public scan alone may not know enough about the operator’s goals to assign the last state confidently.

That makes a useful report partly an evidence plan. It can identify what was inspected, show why an observation might matter and describe what to verify next. It should not pressure a site into building an unused interface solely to improve a generalized score.

A proportionate next step

Run a baseline for a public hostname you operate and read the exact recommendation, including its severity and limitation. Prioritize broken or contradictory signals affecting a real audience before experimenting with optional discovery files. Preserve the original report when you retest so a changed result has a traceable context.

The commercial value of an evidence tool is not the number of warnings it produces. It is whether the operator can make a better, defensible decision. This example shows why a strong assessment can lead to a sensible monitoring decision rather than an unnecessary remediation project.