AI can read most local businesses fine. Almost none give it anything worth quoting.
We drew a random sample of US home- and personal-service businesses across 50 metros and ran all 23 AI-visibility checks on every one. The result was not what we expected, and it is not the result that would sell the most work. Here is the whole dataset, including the parts that argue against our own pitch.
01How this was measured
Method first, because a number is only worth as much as the way it was produced.
Sampling frame. Every business was drawn from Overture Maps (release 2026-08-19.0), a Linux Foundation open dataset merging Meta, Microsoft and TomTom place data. We filtered to US home- and personal-service categories with a listed website and place-confidence above 0.7, took one business per domain so a franchise cannot be counted twice, then drew at random from each of 50 metros with a fixed seed. 2,000 businesses were drawn from an in-scope pool of 173,754.
No score filter. This matters more than anything else on the page. Every business drawn was audited and counted, whatever it scored. We did not drop low performers to make a point, and we did not drop high performers either.
Why that warning is here. We also run a sales pipeline whose database only ever keeps businesses scoring 65 or below. Had we published from it, we would have reported a median in the fifties and called it the state of the market. The real median is 72. A filtered pipeline is not a sample, and any vendor statistic that happens to match the vendor's pitch deserves exactly this question.
The audit. The same 23-check audit we run for clients, unchanged: it fetches the homepage and robots.txt, probes the site again using GPTBot's own user agent to see whether that request is treated differently, and scores 23 weighted checks across four groups. Every check is published in full.
Check it yourself. Every figure on this page is published as machine-readable JSON, and the page is generated from that file rather than typed by hand.
02The finding: access is solved, quotability is not
Each group scored as points earned against points available across the whole sample. One scale, so the four are directly comparable.
This is the whole study in two numbers. Local businesses score 90.7% on AI Crawler Access and 42.2% on Structured Data. An AI engine can almost always reach these sites. What it cannot do is describe them precisely, because the machine-readable facts an engine needs, and the answer-shaped content it quotes, are mostly absent.
The industry story is that small businesses are blocking AI crawlers. In this sample that is mostly false: only 15.4% actually refuse them, 5.7% by disallowing AI bots in robots.txt and 11.4% by refusing a live fetch sent with GPTBot's user agent. The real gap is further down. The site loads, the engine reads it, and there is nothing on the page it can lift into an answer with confidence.
03The distribution
Each bar is a ten-point band. Above each bar is how many businesses landed in it, below is the share of the sample.
The median business scored 72 out of 100 and the mean was 68.3. 55.6% earned an A or a B; 19.6% scored D or F. The tenth percentile was 40 and the ninetieth 88, across a full range of 20 to 100.
If you sell AI-visibility work, this is the inconvenient number: the typical local service business is not a disaster. It is a solid C-plus with one specific hole in it. Any pitch that opens by telling an owner their site is invisible is, for most of them, simply wrong, and they can tell.
04What is actually broken
Share of the 1,354 audited businesses that outright failed each check. Grouped by what the check measures, worst first inside each group.
AI Crawler Access
Structured Data
Content Citability
Entity Signals
The top four failures are all the same failure. 93.1% have no FAQ schema, 93.1% fail FAQPage schema, and 48.0% have no Organization or LocalBusiness schema. Each is a way of saying the site never states, in a form a machine can lift, what this business is, where it works, and what it answers. Meanwhile the basics that everyone worries about, HTTPS and crawler permissions and a title tag, are mostly fine.
05A third of listed businesses have no working website at all
Of 2,000 businesses drawn, 646 (32.3%) could not be audited. These are not excluded quietly; here is exactly why each one failed.
| Reason | Sites | Share |
|---|---|---|
| Domain no longer resolves | 269 | 41.6% |
| Server refused our request | 159 | 24.6% |
| Homepage returns 404 | 79 | 12.2% |
| Other | 68 | 10.5% |
| TLS or certificate failure | 30 | 4.6% |
| No response before timeout | 27 | 4.2% |
| Responded on a later retry | 14 | 2.2% |
The largest single cause is domain no longer resolves, at 41.6% of failures. A business listed today in an open map dataset used by Meta, Microsoft and TomTom has roughly a one-in-three chance of pointing at a website that no longer answers. Those businesses are not scored anywhere on this page, and if they were, every figure above would be worse.
One caveat we will not gloss: "server refused our request" means the site refused our crawler. That is suggestive, not proof, that it refuses GPTBot too.
06By trade
Mean and median score per trade, for trades with at least 15 businesses in the sample. Ordered worst to best.
| Trade | Businesses | Mean | Median |
|---|---|---|---|
| Contractor And Handyman | 137 | 62.7 | 68 |
| Masonry And Chimney | 15 | 65.3 | 68 |
| Pools And Spas | 110 | 66.0 | 70 |
| Pet Services | 95 | 66.2 | 69 |
| Plumbing | 49 | 66.3 | 72 |
| Landscaping | 48 | 66.5 | 70 |
| Auto Detailing And Repair | 444 | 67.5 | 73 |
| Electrical | 44 | 68.0 | 70 |
| Windows And Fencing | 42 | 70.0 | 74 |
| Locksmith | 27 | 70.3 | 72 |
| Remodeling | 56 | 70.3 | 75 |
| Tree Service | 21 | 70.3 | 76 |
| Flooring | 33 | 71.5 | 74 |
| Hvac | 61 | 72.8 | 74 |
| Painting | 27 | 73.6 | 75 |
| Roofing | 54 | 74.3 | 80 |
| Garage And Doors | 32 | 75.3 | 79 |
| Pest Control | 21 | 78.7 | 82 |
Contractor And Handyman came out lowest at a 62.7 mean and Pest Control highest at 78.7. The spread between trades is far smaller than the spread inside any one of them, so trade is a weak predictor. Who built the website matters more than what the business does.
07What this study does not show
The part most vendor research leaves out.
It does not show that a higher score gets you recommended. We hold a second dataset of 1,011 local businesses where we asked Google's Gemini, without search grounding, “Who are the best [trade] in [city]? Name specific local businesses.” Only 13.4% were named at all. Split by audit score, the naming rate did not climb: 9.5% in the 40s, 13.9% in the 50s, 13.8% in the 60s. Within that range the score did not predict being named, and we are not going to pretend otherwise.
Two caveats on that null result. Every business in it scored 65 or below, because that dataset came from a pipeline with a score ceiling, so it contains no high performers and cannot test whether a genuinely good score changes anything. And it is one engine, on one day.
What the score does measure is whether an engine can reach a site, parse what the business is, and quote it accurately. That is a precondition for being recommended, not a guarantee of it. Anyone promising you a citation in ChatGPT is guessing.
08Limitations
- Overture's coverage is very good but not complete, and its category labels are imperfect. A mislabelled business lands in the wrong trade.
- Only businesses with a website in the dataset could be sampled. Those trading purely off a Facebook page or a Google listing were excluded, and they are almost certainly less visible to AI than anything counted here.
- The audit reads the homepage plus robots.txt. A site with good schema on interior pages but none on the homepage scores lower than a full crawl would give it.
- Scores are a snapshot taken on 2026-09-12. Sites change.
- US metros only. Nothing here should be read as applying to other countries.
09Reuse this
Free to cite and republish with attribution to NamedHQ, under CC BY 4.0. The underlying numbers are at /ai-visibility-index.json. If you want a cut of the data we have not published, ask us.
10The same data, by trade
Every trade with a large enough sample has its own page: its real median, the checks it fails most, and where it ranks against the others.
Where does your site sit on this chart?
Run the same 23 checks on your own site. No email, no call, no card: the score and every failing check appear on the page.
Scan my site free