More people now ask an AI assistant how to file a tax extension, replace a Social Security card, or apply for student aid than read the agency's own instructions page. We ran our own AgentReady audit over 14 major US government and public-service websites on 2026-07-22 to see how ready the public sector is to be read and cited correctly by one at the moment a citizen asks.

51.4/100 average score among the 11 sites that let our fetcher in. Median 50. That is the lowest average of any industry we have benchmarked. But the sharper finding is which three sites did not answer at all.

The real finding: the three services people most need help with would not answer

3 of the 14 sites — irs.gov, ssa.gov, and studentaid.gov — returned no response to a single, clearly-identified, respectful server-side request. The IRS and Social Security sites failed the connection outright, and the federal student-aid site timed out past a 12-second budget on two separate attempts. Filing taxes, claiming Social Security, applying for college aid: the three government tasks Americans most often ask an assistant for help with are the three whose front doors an assistant cannot get through. A closed door is not a low score, it is no data at all, so these three are excluded from the average rather than penalized, our standard honest-scope rule. But the exclusion is the story. Among the 11 sites that did answer, not one had published an llms.txt file, and not one had llms-full.txt or agents.md either — the machine-readable files AI assistants look for first. Public services paid for by everyone are, on this measure, harder for an AI to read on a citizen's behalf than the average private business.

The ranking

#	Site	Score
1	usa.gov	70
2	medicare.gov	70
3	nasa.gov	70
4	texas.gov	70
5	va.gov	50
6	login.gov	50
7	ny.gov	50
8	healthcare.gov	45
9	cdc.gov	35
10	sec.gov	35
11	weather.gov	20

3 sites attempted (irs.gov, ssa.gov, studentaid.gov) returned no auditable homepage to our fetcher — a reproducible connection failure or past-12-second timeout — and are excluded rather than scored as 0, the same honest-scope rule as every benchmark on this site. Note weather.gov, the site tens of millions check daily, scored lowest of those that answered at 20, and both cdc.gov and sec.gov, the public's primary sources for health guidance and market filings, landed at 35.

Methodology

One server-side fetch of each public homepage, clear identifying User-Agent, 9 machine-readability checks (structured data, AI-crawler access, answer-shaped content, metadata, headings, sitemap, no-JS legibility, entity clarity, crawlability), weighted 0-100. Homepage-only snapshot, not a full-site audit, dated 2026-07-22. Public data only, no personal data, one request per site, and no attempt to work around any bot protection. This is a factual machine-readability scan, not a claim about Section 508, accessibility law, or any agency's compliance obligations. Scores change as sites change. Æ Studio is AI-operated and human-reviewed, and says so. Same audit engine and weights as the SaaS edition, the Insurance edition, the Real Estate edition, the Financial Services edition, the Healthcare Systems edition, and — run on the AI industry's own homepages — the AI-companies edition. A separate, related scan: the Web Accessibility edition (real axe-core WCAG 2.2 scan, not this AI-visibility method). Every industry we have scored, in one index: The State of Machine-Readable Business, 2026. Full weights and rationale: the published methodology.