The State of Machine-Readable Business, 2026

The systems that recommend businesses have changed. A customer used to open ten blue links and choose. Now an AI reads the web on their behalf and hands back a short list, and most business owners have no idea whether their own site made that list or was skipped because a machine could not read it. This is the index of what we found when we checked, one rubric, every industry we have run so far, dated and citable.

64.9/100 average AI-visibility score across 158 homepages in fifteen industries, on one nine-point machine-readability rubric. The gap between the leaders and the laggards is not size or budget. It is whether anyone did the small, unglamorous work of making the site legible to an agent.

Every benchmark, one table

Each row below is a separate audit this studio ran in 2026, on the identical nine-point test. The scores are real, the sites are named on each detail page, and the method is public.

Weighted across all 158 homepages, the average lands at 64.9/100. SaaS and e-commerce lead because those teams already treat structured data as table stakes. Regulated, relationship-driven, or brand-heritage fields — dental, financial advice, legal, hospitality — trail by ten to twenty points, not because the work is harder but because almost no one in those fields has started. Real estate and hospitality trail furthest, and both stand apart for the same reason: real estate had 4 of 14 brokerages and portals block our fetcher outright, and hospitality was worse still — 10 of 14 major hotel brands blocked us, leaving only 4 sites scorable at all, the highest block rate of any industry benchmarked so far. Automotive shows the same split at brand level: Hyundai and Kia led while Tesla, BMW, Mercedes-Benz, Audi, GM and Honda all blocked the fetcher outright. Insurance carriers came next after hotels for hostility: 6 of 14 — Geico, Allstate, Farmers, Nationwide, The Hartford and USAA — would not return a homepage to a plainly-identified request, a 43% block rate, even as AI shopping assistants become how people compare coverage. Government and public-service sites sit lowest of all at 51.4, and carry their own version of the same pattern: irs.gov, ssa.gov and studentaid.gov — taxes, Social Security and student aid, the three services people most often ask an assistant to help with — each returned no response to a plainly-identified request, and not one of the 11 scorable public sites had published an llms.txt. Banking lands at 57.3, with the same one-in-eleven pattern on llms.txt: only usbank.com publishes one, and 3 of 14 banks attempted, including TD Bank and KeyBank, blocked our fetcher outright. Airlines is now the lowest average of all at 43.6: 7 of 14 airlines attempted, including United, Delta, American, and JetBlue, blocked our fetcher outright, a 50% block rate, and of the 7 that answered only Sun Country publishes an llms.txt. Telecom breaks the pattern the other direction: the highest average of any edition at 76.3, and the highest llms.txt adoption rate measured (5 of 8 scorable carriers) — yet T-Mobile, one of the Big Three, still blocked our fetcher outright alongside 5 smaller providers.

The one finding that repeats everywhere

Across every industry, the basics of old-fashioned SEO are handled. Robots.txt and sitemaps are close to universal. The layer that AI agents actually read is where the sites fall down, and they fall down in the same order every time.

• agents.md is almost never present. Financial advisors: 13% published one. Dental: 3 of 13. Real estate: 2 of 10 scorable sites. Legal: zero of 11. This is the file that tells an AI agent, in plain language, what a business does and who it serves. Nearly everyone leaves it blank.
• llms.txt is a coin flip at best. 44% of financial firms, 23% of dental practices, 20% of real-estate sites, and roughly one in five SaaS sites publish it. E-commerce does better at 56%, still barely half. Cleveland Clinic was the only one of 10 scorable hospital systems to publish one at all.
• Structured product and service data is the expensive miss. Only 56% of e-commerce brands publish the schema a shopping agent needs to compare and recommend them. The half without it are invisible at the exact moment a customer's agent is deciding what to buy.

Brand size does not protect you

The sharpest result in the whole dataset was not an average, it was a spread. In the financial-advisor benchmark, three of the most recognized names in American finance scored near the bottom: Northwestern Mutual at 15, Charles Schwab at 25, Vanguard at 50. Newer digital-native firms like Betterment and Hightower cleared 90. The real-estate benchmark repeats the same pattern: Sotheby's International Realty (15) and Century 21 (20) scored worst, while Opendoor (95) and Compass (85), both younger, tech-first companies, led the field. Automotive and healthcare both go further still: Tesla, BMW, Mercedes-Benz, Audi, GM and Honda, and separately Mayo Clinic and Johns Hopkins Medicine, are among the most recognized brand names in their industries and every one of them blocked our fetcher outright — not a low score, no data at all. A household name earns trust with humans. It earns nothing with an agent that has never heard of anyone, reads only what the page makes readable, and can't even get past the front door of some of the biggest names in the room.

A companion track: can assistive technology read the page at all?

Machine-readability for AI and machine-readability for a human using a screen reader are close cousins. So we ran a second, separate benchmark on a different rubric: a real WCAG 2.2 accessibility scan of ten major US retail homepages.

58.3/100 average accessibility score, median 56.5. Walmart scored a perfect 100. Kohl's scored 6, with two critical failures including buttons that have no name a screen reader can announce. Eight of the ten failed a basic landmark-region check.

Methodology and honest scope

Every AI-visibility score here uses the same rubric: one server-side fetch of each public homepage, a clearly identifying User-Agent, nine weighted machine-readability checks, public data only, no personal data, one request per site. These are homepage snapshots, not full-site audits, each dated on its own detail page. Full weights and rationale: the published methodology. The accessibility track is a separate rubric, a severity-weighted axe-core WCAG 2.2 scan run in a real browser, weights described on that benchmark. Some sites in each run returned no auditable result to our fetcher because of bot protection, which is itself an AI-visibility risk and is noted on each detail page rather than scored as a pass. Numbers were true on the date each benchmark ran. We update this index as we run new industries.

See also: we ran this same audit on the AI industry's own homepages -- 30% blocked our fetcher outright.

Built by Æ Studio, AI-operated and human-reviewed. Contact: alexander.k.eliot@gmail.com