Most business websites are not hiding from AI. They are unreadable to it.
We scored 252 independent Northwest Ohio business websites with the same check the practice publishes for anyone to use, and the pattern held across every sector we looked at. These businesses are reachable. Crawlers get in, pages return, robots.txt lets them through. Where they fall short, in every one of the ten sectors, is telling a machine who the business is.
More than one in four could not say what the business does. One in nine could not state its own name in a form a machine could read back. Four in ten published no structured data at all.
What we checked
The instrument is the practice's AI visibility check, and its scoring model is published in full. It fetches a site the way an AI crawler does: the homepage, robots.txt, the sitemap, llms.txt, the favicon, and the homepage a second time under a published AI crawler's user agent. Eight requests, no recursion, no browser spoofing. If a site challenges the declared agent, that is itself a finding.
The check scores 100 points across four sections: whether AI can reach the site, whether it can read it, whether it can say who the business is, and basic hygiene. It never queries a live model. It measures what a crawler can retrieve, which is the input to every answer an assistant gives about a business.
The sample is the top search results for a set of sector queries across the Toledo metro, taking organic and local-pack entries, minus a published stoplist. We removed national chains and franchise landing pages, because those measure a national brand's infrastructure rather than a local business's. We removed platforms, directories, review sites, government pages, trade bodies, and institutions for the same reason. What remained was 259 independent local businesses across ten sectors. Seven could not be reached at all, which leaves 252 scored.1
Every site was checked once, on 31 August 2026. Sites change. These are measurements, not verdicts, and no business is named anywhere in this study.
What we found
The median score was 80 out of 100, and half the sample landed in the top band.
The scores hide the shape of the problem. What a machine reading these sites could not determine:
| The site does not make readable | Share of 252 |
|---|---|
| Anything corroborating the business exists elsewhere | 77.8% |
| Where the business is | 44.0% |
| What the business does | 27.4% |
| How to reach it | 21.0% |
| The name of the business | 11.1% |
The most-failed signals tell the same story from the other direction. Structured data absent entirely on 43.3 percent. No linked profiles anywhere on 77.8 percent. No working headline describing the business on 35.7 percent. An incomplete business entity, missing some combination of name, address, phone or hours, on 58.3 percent.
Access was rarely the problem. Only 17.9 percent blocked an AI agent at the firewall while allowing it in robots.txt. The check cannot see why a block happens; a security product set to stop bots is a common cause. What that block looks like from the owner’s side, and how to tell whether a given site is in that group, is the subject of the Field Note on whether a website is blocking AI.
What the sectors show
| Sector | Sites | Median |
|---|---|---|
| Education and childcare | 18 | 71 |
| Funeral homes | 23 | 72 |
| Manufacturing and print | 22 | 75 |
| Hospitality and catering | 20 | 76 |
| Professional services | 15 | 78 |
| Construction and home building | 16 | 80 |
| Retail and personal care | 22 | 80 |
| Automotive | 18 | 84 |
| Healthcare | 31 | 84 |
| Trades | 67 | 84 |
The trades sit at the top. Plumbers, electricians, roofers and HVAC contractors are more legible to machines than schools, funeral homes and professional firms.
That inverts the usual assumption. The likely reason is local search: the trades have been competing in local search for fifteen years. Somebody sold them structured data and a Google Business Profile because their customers were plainly searching. Professional firms were never under that pressure, because their work arrived by referral. The referral engine worked, so nobody rebuilt the storefront. The storefront is now what the machine reads.
Why this is a legibility problem
An assistant answering “who should I call for this” is not browsing. It is assembling an answer from what it can parse. A site that loads beautifully for a person and says nothing machine-readable about the business is, to that assistant, a blank page with a nice photograph on it.
The failures cluster in exactly the places a business would never think to look. Nobody proofreads their structured data. Nobody notices that their site never states its own city in a form a machine can extract. They are neither design problems nor content problems. They live in the seam between the two, and the seam belongs to nobody.
The three cheapest fixes
- Publish structured data that says who you are.
- Name, what you do, address, phone, hours. This was the largest single gap in the sample and it is among the smallest jobs on the list.
- Add your linked profiles.
- The sameAs property connects a site to the business's other identities, and 77.8 percent of these sites have none. It costs a few lines.
- Say what you do in your headline and your title.
- The brand name alone is not enough. Say what the business actually does, in the words a person would type.
None of it requires a redesign.
Follow-on reads
The practice has since cut the same sample in more ways, each reported in aggregate like the rest. These are the figures its service pages cite.
| Measure | Result |
|---|---|
| Share of each section’s points the average site earned: reach, reading, who the business is, housekeeping | 90.8% · 82.2% · 51.7% · 54.0% |
| Sites whose site icon does not load | 57.1% |
| Sites without a complete card for when a page is shared | 29.4% |
| Sites without an XML sitemap | 27.4% |
| Of the 204 domains that receive email: SPF published · DMARC published · DMARC set to enforce (public DNS, read 26 September 2026; DKIM not counted) | 87.3% · 48.0% · 17.2% |
| Of the 213 homepages that could be fetched: a Google tag · a Meta pixel · neither (26 September 2026; site-builder pages rendered with tracking blocked to catch late-loading tags) | 68.1% · 13.1% · 30.5% |
Limits
This is one metro, one snapshot, one instrument. The scoring is deterministic and never queries a live model, so it measures what a crawler can retrieve rather than what any particular assistant will say. A site can score well and still be described badly, and a site can score badly and still be well known through other sources.
And a disclosure worth making plainly: the practice built the instrument. Our own site was excluded from the sample along with every site we manage, before the run, by a list. Weigh that as you like. The method is published so the sample can be rebuilt and the study re-run by anyone who wants to check us, and the check itself is free, ungated, and does not ask for an email.
Binary Glyph is a brand and marketing practice in Toledo, Ohio. If you want to see what a machine reads on your own site, the check used in this study is free and ungated, and the reference on how AI describes a business explains what moves the answer. If the gap turns out to be larger than three fixes, begin a conversation →
Measured on 31 August 2026, with the follow-on reads dated where they appear. These figures describe those days.