Is Your Website Turning AI Readers Away? We Checked 200 Dealers
We fetched 200 independent dealer homepages as a browser and then as the four readers the AI companies send out. One in eight refused at least one, and the pattern says it was a plugin default, not an owner's decision.
On September 17, 2026 we fetched the homepage of 200 independent alarm and security dealer websites. First as an ordinary web browser, then as each of the four readers the AI companies send out.
On 168 of those 200 the browser fetch came back clean, so we could tell what happened next. Twenty-one of the 168 - 12.5% - refused at least one AI reader. Five of them, 3.0%, refused the readers that decide whether you get named in an answer.
Almost none of those owners chose it.
The other 32 sites we are not counting either way. Something in front of the site challenged our request before we ever got a clean read, so the honest verdict there is that we could not tell. Every percentage below is out of the 168 we could actually observe.
There are two kinds of AI reader, and they do different jobs
This is the part almost nobody gets right, and it changes what the number means.
Training crawlers are GPTBot, ClaudeBot, CCBot, Google-Extended and a few others. They collect pages to train a model. A site that turns them away keeps its own words out of AI training. That is all it does. An assistant can still open that same site later to answer a live question.
Answer readers are OAI-SearchBot, ChatGPT-User, PerplexityBot and their siblings. These are what an assistant sends out in the moment. A homeowner asks who installs alarm systems in their area, the assistant goes and reads a handful of sites, and then it names a few companies. Turning these away is the expensive one.
Of the 21 dealer sites that refused something, 16 refused training crawlers only, and 5 refused an answer reader.
Refusing the training crawlers on purpose is a fair choice. Plenty of owners would rather their words stayed out of a model nobody asked them about. The cost is real but small: you give up the chance that your own explanation of your work becomes part of what the model knows, and live lookups still go through. Refusing an answer reader is the one that costs you the call.
It clusters by platform, and that is the whole finding
| What the site runs on | Dealer sites we could check | Refused an AI reader | Rate |
|---|---|---|---|
| Self-hosted WordPress | 86 | 17 | 19.8% |
| Sitting behind Cloudflare | 15 | 3 | 20.0% |
| Wix | 7 | 0 | 0.0% |
| Squarespace | 8 | 0 | 0.0% |
| Plain Apache or nginx | 19 | 0 | 0.0% |
| Everything else | 33 | 1 | 3.0% |
| All observable dealer sites | 168 | 21 | 12.5% |
Wix, Squarespace, plain Apache and plain nginx refused nothing at all. Not one site out of 34. Self-hosted WordPress and sites sitting behind Cloudflare refused at about one in five.
We ran the identical scan on 200 dentist and chiropractor websites the same day, to see whether this was a security-industry habit. It is not. Of 173 observable practice sites, 13.3% refused something and 5.8% refused an answer reader. Same shape, same two platforms.
Nobody clicked a box that said "keep ChatGPT out"
The clustering is the tell. If owners were making a decision about AI, the refusals would be scattered across every platform about evenly. They are not. They land almost entirely on the two setups that ship with bot blocking already switched on.
A security plugin on a self-hosted WordPress site arrives with a list of user agents it treats as bad robots. A firewall in front of a site ships with managed rules that do the same thing. Both lists were written to stop scrapers and spam, most of them before any of these readers existed. The readers OpenAI, Anthropic and Perplexity send out look like scrapers to a rule from five years ago.
So the site works fine, the owner has no idea, and the refusal was installed by a default nobody reviewed.
Reading your robots.txt file will not find this
We pulled robots.txt on all 400 domains in the study. 338 came back readable. Of those, 6 - 1.8% - name an answer reader and tell it not to read the site. At the server, 4.4% refused one.
The file under-reports by about two and a half times, because the refusal is usually not in the file. It is in the plugin or the firewall sitting in front of the file. Only 43 of the 338 readable files mention an AI reader at all.
Any audit that grades a site's AI access from robots.txt alone is reading the wrong instrument.
Why this is worth ten minutes now
People who asked an AI assistant to recommend a local business went from 6% to 45% in one year (BrightLocal 2026, n=1,002). That is why this moved from a technical curiosity to something worth ten minutes.
If you are one of the 87.5% whose site reads clean, this costs you nothing. If you are in the 12.5%, find out which kind of refusal it is before you do anything about it.
How to check your own site in about ten seconds
We built a free checker that does exactly what the study did: aisecurityedge.com/tools/ai-crawler-check. Type your address, and it tells you, reader by reader, which ones your server let in and which ones it refused. No email required, nothing to install, and it never reports a refusal it did not observe.
If it comes back saying an answer reader was refused, send that result to whoever manages your website or your firewall. The fix is usually one allow rule, not a rebuild.
Frequently Asked Questions
Does turning away GPTBot actually hurt my business?
Not in the way most people assume. GPTBot collects pages to train a model, so refusing it keeps your words out of that training. It does not stop ChatGPT from opening your site to answer a live question, because that is a different reader with a different name. If a training-crawler refusal is all your site is doing, treat it as the smaller thing it is.
How do I tell which kind of reader my site is refusing?
By name. Ask whoever runs your site, or check your firewall dashboard, for the list of user agents being turned away. GPTBot, ClaudeBot, CCBot, Google-Extended and Applebot-Extended are training crawlers. OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-SearchBot and Claude-User are answer readers. The free checker above does the same thing without you having to ask anybody.
My web company says everything is fine. Who is right?
Both of you, probably. Their test is a browser loading the page, and the page loads. The refusal only shows up when the request identifies itself as an AI reader, which nothing in a normal site check ever does. Show them the reader-by-reader result instead.
If I fix this, will ChatGPT start recommending my company?
No, and anyone who tells you otherwise is selling something. Letting the readers in is necessary, not sufficient. 87.5% of the dealer sites we checked already let every reader in, and most still do not get named when a homeowner asks. Being readable is the floor, not the win.
---
AI Security Edge works on one thing for alarm dealers: getting named by ChatGPT and Google when a homeowner asks who to hire. It is run by Thad Paschall, who started Protect America and created over $600 million in revenue.