A small study from September 2026, written for business owners — not for technicians.
When someone asks ChatGPT or Perplexity to recommend a cabinet maker, an accountant, or an insurance broker in Winnipeg, the assistant doesn’t already know the answer. It goes and reads websites first — quickly, quietly, in the background — and then names the businesses whose websites it was able to read.
If your website won’t let it in, you are not in the answer. Not near the bottom. Not there at all.
We wanted to know how often that happens to ordinary, well-established businesses. So we checked.
This is a first look — 48 websites, one question. A larger, more comprehensive study, published as a white paper, follows later this month.
What we did
Between 4 and 8 September 2026 we tested 48 business websites — 26 in Manitoba and 22 across the northern United States. They weren’t chosen at random; they are the family-owned and long-established businesses we were already looking at — manufacturers, insurance brokers, accountants, food shops, retailers and trades. Most have been in business for decades.
For each website we knocked on the front door three times: once as an ordinary person using a web browser, and twice as the reading programs that ChatGPT and Perplexity send out. Then we compared the answers.
We also read each website’s welcome sign — the small public notice every website carries that says which visitors are allowed in. (If you’ve never heard of it, that’s normal. Your web designer has.)
A website counted as turning AI away if it opened the door for the ordinary visitor but refused the AI programs.
What we found
| Websites tested | Turned AI away | That’s about | |
|---|---|---|---|
| Manitoba | 26 | 7 | 1 in 4 |
| Northern US | 22 | 1 | 1 in 20 |
About one in four of the Manitoba websites we tested turns at least one of the two AI assistants away. Among the American ones, about one in twenty. Six times the rate.
Here is the part that matters most. Every one of the seven Manitoba websites that turned AI away had a welcome sign saying “everyone is welcome” — or no sign at all. The sign says come in. The door says no. An owner who checked the sign would believe their website is open to AI. It isn’t.
We ran the whole test again a few hours later to see what held up. Five of the seven were exactly the same the second time. Two we couldn’t re-test — one website wasn’t answering anyone at that moment, and one had started turning away every visitor, ordinary people included, so it no longer counted. So the number we’re comfortable standing behind is a range: at least one in five, confirmed twice; one in four on the first pass. Either way, it’s several times what we saw south of the border.
What we got wrong first
Our very first look, on 4 September, covered fifteen websites and suggested that half of the Manitoba ones were turning AI away. We decided then that we wouldn’t say that number to anyone until we’d checked about fifty. At 48, the honest number is closer to one in four. Small samples exaggerate. We’re publishing the corrected number, not the exciting one.
The explanation we tested — and dropped
Our first theory was that it came down to which company hosts the website — that a few Canadian hosting companies were switching this on for everyone. We had a reason to think so: earlier this year our own website was quietly turning away these same AI programs for fifty days, because of a setting at the hosting company that we couldn’t see and that only a support request could fix.
So we checked where the seven Manitoba websites are hosted. They’re spread across six different companies. And two of those companies also host websites in our sample that let AI in without any trouble. Same host, opposite result.
That rules the hosting companies out as the whole story. What fits better is that on each website, somewhere, a protective setting has been switched on — by a security tool, a hosting company, or someone who set the site up years ago and meant well. We haven’t looked inside anyone’s settings — they aren’t ours to look at — so that’s our best explanation, not a proven fact. But if it’s right, it’s good news: the fix isn’t a fight with a hosting company. It’s a switch.
What this is, and what it isn’t
It is a real test of 48 real websites, done twice. As far as we know, it’s the first number of its kind for Manitoba.
It isn’t a survey of all Manitoba businesses, and we don’t claim it is. It also isn’t perfect: some websites check a visitor’s identity more carefully than our test could, and might let the real AI programs in while turning our test away. We can’t tell those apart from the outside. That is one more reason we’re calling this a first look rather than a final word — and why it isn’t a verdict on any one website. That takes a proper look at that website.
What it means for you
If your business has been around for years and you’ve noticed that AI assistants never seem to mention you, this may be part of the reason. It isn’t that your reputation isn’t real. It’s that the assistant may never have been let in to read about it.
Being let in is only the first thing a machine checks. What it reads next — whether your site says plainly what you do, where you do it, and who you are — is the subject of our plain-language guide to schema markup, the labels that tell a machine who you are.
The businesses in our sample have done nothing wrong. Most of them will never know unless someone tells them.
How to find out about yours
Ask whoever looks after your website one question: “Does our site let ChatGPT’s and Perplexity’s reading programs in — and can you test it, rather than just check the setting?” The second half of that question matters. The welcome sign can say yes while the door says no.
Or ask us. Checking this — along with the twenty other things an AI assistant reads before it decides whether to recommend a business — is what our Foundation Audit is. There’s no obligation, and you keep the written summary either way.
If your business is in Winnipeg, this check is where our work begins. It is the first of the five levers described on our SEO in Winnipeg page: measure first, then fix what the measurement shows.
What comes next. This first look answers one question for 48 websites. Later this month we’ll publish a larger, more comprehensive study as a white paper — more businesses, more of the northern corridor, and the other things a machine looks at before it recommends anyone. If you’d like a copy when it’s ready, write to bespoke@obrecs.com and we’ll send it.
Measured by OBrecs, Winnipeg, 4–8 September 2026. 48 websites (26 Manitoba, 22 northern US), each tested twice on 8 September. No business was contacted, and no business is named.