What AI-recommended local businesses' websites have in common
In our local-business study we asked ChatGPT, Claude and Gemini for plumbers, dentists, accountants and other local businesses in 16 cities. Many answers linked to the website of a business they recommended. So we ran our own checks on those websites: can AI crawlers read them, do they tell search engines what the business is, are their basics in order?
The short version: the recommended businesses’ websites share the basics that let an assistant read them, and not much else.
- None of them blocked an AI search crawler in robots.txt. Not one of the 485 sites we could read blocks OAI-SearchBot, Claude-SearchBot or PerplexityBot there. Among the 1,000 most-visited websites, 18% of those with a readable robots.txt do.
- Almost all have the plain basics: 96% are built for phones, 95% forward http:// to https://, 93% show the current year in the footer.
- Many skip things often recommended for being found: only 53% describe themselves as a local business in structured data (schema), and 23% have no business structured data at all. 32% have an llms.txt file.
- Email security is weak: only 20% enforce DMARC, the setting that stops others sending email in the business’s name.
- Businesses named by two or three assistants look almost the same as those named by one. Nothing we measured separates them clearly.
The data is in a CSV file, one row per website.
What we checked
We started from the 432 answers of our local-business study: six trades (plumber, electrician, dentist, accountant, hair salon, wedding photographer) in 16 cities across the United States, United Kingdom, Canada, Australia, Ireland, New Zealand, Singapore and South Africa, asked on 19 November 2026.
From those answers we took every cited website that belongs to a business the same answer named, matching the web address to the business’s name. We reviewed the matches by hand and removed 8 that weren’t a named business’s own site (for example a city council site and a booking platform). That left 570 business websites. We then counted how many of the three assistants named each business in the same city and trade, in any of the study’s runs:
| Named by | Websites |
|---|---|
| One assistant | 356 |
| Two assistants | 168 |
| All three | 46 |
On 19 November 2026 we checked each website once, with the same code as our free tools: the AI crawler check, the website check and the site health check, which includes the email security check. For 485 of the 570 sites our website check could read the home page. The other 85 refused our check (33 answered “forbidden”, HTTP 403), returned an error or didn’t answer in time. The figures below are for those 485.
What they have in common
| Check | Share of the 485 sites |
|---|---|
| Built for phones (viewport set) | 96% |
| Forwards http:// to https:// | 95% |
| Current year in the footer | 93% |
| ChatGPT’s crawlers can read the home page | 92% |
| Claude’s crawlers can read the home page | 93% |
| Googlebot can read the home page | 90% |
| Perplexity’s crawlers can read the home page | 88% |
| All four can read it | 84% |
| Has a robots.txt file | 90% |
| Scores 80 or more on our website check (median 93) | 89% |
No robots.txt blocks AI search. Not one of the 485 sites blocks an AI search crawler in robots.txt. 15 sites (3%) block a crawler that collects training data (GPTBot, ClaudeBot, Google-Extended or CCBot), which doesn’t affect search answers according to the companies’ documentation.
The blocks we saw come from servers, not robots.txt. On 30 sites (6%) at least one assistant’s crawler was refused or challenged by the server, or the page asked not to be indexed. Perplexity was refused most often (24 sites, 18 of them behind Cloudflare), ChatGPT and Claude on a handful each. Four home pages carried a “noindex” tag, telling every search engine and assistant not to list them. The assistants recommended these businesses anyway: an assistant can learn about a business from directories and reviews without reading its site.
On another 48 sites (10%) we couldn’t tell for at least one assistant: some sites treat a test from our server differently from the real crawlers, which the companies can recognise by their network addresses.
Where they differ
| Check | Share of the 485 sites |
|---|---|
| Structured data saying it’s a local business (LocalBusiness or a type like Dentist, Plumber) | 53% |
| Only general Organization structured data | 24% |
| No business structured data | 23% |
| Page title of a sensible length (10 to 65 characters) | 78% |
| Meta description of a sensible length | 72% |
| Phone number on the home or contact page | 77% |
| Street address on the home or contact page | 70% |
| Exactly one main heading (H1) | 69% |
| Has an llms.txt file | 32% |
| Sends the HSTS security header | 49% |
| Security-header grade A or B | 14% |
| Security-header grade F | 60% |
| Has an SPF email record | 84% |
| Has a DMARC record | 62% |
| DMARC enforced (quarantine or reject) | 20% |
Structured data, page descriptions, an llms.txt file and email security vary a lot among businesses the assistants recommended. So none of them is required to be recommended. That doesn’t mean they don’t help: we can’t see the businesses that weren’t named (see the limits below).
Across all 570 domains, 37% have no DMARC record and 44% have one that only monitors (p=none). Only 19% (111) tell receiving servers to reject or quarantine email that fakes the domain. That has nothing to do with AI, but it’s an easy fix that protects a business’s customers from fake invoices sent in its name.
Named by one assistant or by several
If some website feature helped businesses get named, we’d expect businesses named by several assistants to have it more often. They mostly don’t:
| Check | Named by one (306 sites) | Named by two or three (179) | Named by all three (37) |
|---|---|---|---|
| All four AI crawlers can read it | 84% | 84% | 92% |
| LocalBusiness-type structured data | 54% | 53% | 59% |
| No business structured data | 23% | 23% | 19% |
| Meta description of a sensible length | 72% | 72% | 81% |
| Phone number on the page | 75% | 81% | 86% |
| Forwards http:// to https:// | 93% | 98% | 100% |
| Has an llms.txt file | 30% | 35% | 43% |
| DMARC enforced | 19% | 20% | 22% |
| Median website check score | 92.5 | 93 | 94 |
The businesses named by all three assistants do a little better on several checks, but there are only 37 of them, and gaps of a few points between groups this size are within what chance gives. The gap least likely to be chance is http:// forwarding to https://, 93% against 98%, and with more than 30 checks compared even a gap like that can turn up by luck. We read this table as “no clear difference”, not as “llms.txt gets you recommended”.
By trade and country
| Trade | Sites | LocalBusiness-type data | DMARC enforced | Description ok | Median score |
|---|---|---|---|---|---|
| Accountant | 123 | 46% | 39% | 77% | 93 |
| Dentist | 84 | 60% | 19% | 80% | 96 |
| Electrician | 66 | 65% | 18% | 73% | 93 |
| Hair salon | 85 | 48% | 9% | 55% | 91 |
| Plumber | 52 | 56% | 13% | 77% | 94 |
| Wedding photographer | 75 | 53% | 5% | 69% | 86 |
Accountants enforce DMARC far more often than the other trades (39%); wedding photographers and hair salons rarely do. Hair salons most often lack a usable page description (45% don’t have one).
| Country | Sites | LocalBusiness-type data | Any business data | DMARC enforced | Median score |
|---|---|---|---|---|---|
| United States | 134 | 60% | 81% | 25% | 92 |
| United Kingdom | 102 | 48% | 80% | 21% | 93.5 |
| Australia | 71 | 56% | 82% | 13% | 93 |
| Canada | 63 | 52% | 67% | 14% | 93 |
| New Zealand | 37 | 46% | 76% | 14% | 94 |
| Ireland | 33 | 61% | 79% | 15% | 96 |
| Singapore | 25 | 48% | 72% | 32% | 93 |
| South Africa | 22 | 36% | 64% | 27% | 91.5 |
The groups for Singapore and South Africa are small; read their figures loosely.
What this means for a small business
- Don’t lock the assistants out. None of the recommended sites blocked AI search in robots.txt, but some security settings refuse AI crawlers without the owner knowing (Perplexity most often, mostly behind Cloudflare). Our free AI crawler check shows in a minute what each assistant’s crawler gets from your site.
- Get the basics right first: a phone-friendly site on HTTPS, your phone number and address on the page, a clear title and description. Most recommended businesses had them. Our site health check covers HTTPS and the technical side.
- Add business structured data. Half the recommended businesses didn’t describe themselves as a local business in structured data, so it isn’t required, but it’s how your name, address and phone reach search engines in their own format. Our business schema generator writes it for you.
- Fix your email security anyway. Four in five recommended businesses don’t enforce DMARC. It won’t change AI answers, but it stops others sending email in your name. The email security check tells you what to add.
- Then check the answers themselves. A good website is the floor, not the reason an assistant picks you: in our earlier study the assistants leaned heavily on directories and review sites. Our AI answer check shows whether Gemini names your business.
How we counted, and the limits
- No comparison group of businesses that weren’t named. The assistants’ answers don’t list the businesses they passed over, so we can’t say whether recommended businesses’ websites are better than average. A feature that almost all recommended sites have (like HTTPS) may be just as common among sites that weren’t recommended. This study shows what recommended sites look like, not what causes a recommendation.
- Only businesses whose own site was cited. A business enters the sample only when an answer linked to its website and the web address matches its name. Businesses named without a link, or with a web address unlike their name, are left out, and so are the ones whose site our check couldn’t read (85 of 570). Some well-known chains and large firms are included (for example national accountancy firms and salon chains).
- One check per site, from our server, on one day. The AI crawler check sends requests that name each crawler from our own server; some sites treat the real crawlers differently, and we report those as “couldn’t tell” rather than guessing. Blocks by a server can change from day to day.
- Our checks, not the assistants’ criteria. Nobody outside OpenAI, Anthropic and Google knows exactly how their assistants choose businesses. We measured what a site owner can control and see.
- “Named by” counts the three assistants across all runs of the earlier study, matched by web address and name, so a few businesses may be counted under the wrong group.
The code that ran the checks is part of the OOO Works visibility engine and uses the same checks as our free tools. If you use the data, please link to this page.