Who ChatGPT Recommends in Toronto: We Asked 120 Questions Across Eight Industries

What ChatGPT recommended across 120 questions

Ask ChatGPT where an 84-year-old mother who can no longer live alone should move in Mississauga and you get a short list. In our test the typical retirement home answer linked to fewer than three homes. Amica's website turned up in six of 15 answers and Chartwell's in five. The regulator's public register appeared in nine.

Ask it for a wedding venue for 150 guests in Toronto and the answer runs longer and almost never repeats itself. Across 15 venue questions ChatGPT linked to 68 different venues, and nine in ten of them showed up exactly once.

That split is the main thing a small study we ran on the afternoon of 27 October 2026 turned up. We wrote 15 questions for each of eight industries, phrased the way a resident of the Greater Toronto Area would ask them, put each one to ChatGPT in a fresh logged-out chat, and logged the websites every answer linked to. Care categories, where the decision is about a parent or a pregnancy, produced short lists that leaned on the same operators and on public registers. Taste categories, such as schools, venues and camps, produced long lists that scattered.

For a business owner the two patterns call for different responses. Where the list is short and repeats, being left off means being left off most of the time. Where it scatters, whether you are named depends on how the question is worded, and no single answer tells you much.

"Businesses linked per answer" leaves out directories, regulators, government sites and review sites, which we counted separately. "Named only once" is the share of different businesses that appeared in a single answer. Home care has 14 answers because one question returned no usable reply. The method and its limits are further down, and every question is listed at the end.

Retirement homes: a short list, led by Amica and Chartwell

Retirement homes produced the shortest answers of the eight industries: 42 links to homes across 15 answers, or 2.8 per answer. Six operators with residences across the region, Amica, Chartwell, Sunrise, Cogir, Aspira and Verve, accounted for 20 of those 42 links, or 48%. Amica led with six answers and Chartwell followed with five. Sunrise and V!VA appeared in three each, and Bradgate Arms, Cogir, Aspira and Delmanor in two.

The rest of the region barely registered. ChatGPT linked 25 different homes in total, and 17 of them appeared once. The questions covered 13 places from Burlington to Pickering, so a home that is the obvious answer in its own town could still be absent from 14 of 15 replies.

What ChatGPT did more often than name a second or third home was send the reader to a public source. The Retirement Homes Regulatory Authority, which licenses Ontario retirement homes and publishes a register of them, was linked in nine answers. Ontario Health atHome, Comfort Life, Peel Region and a regional health service finder added five more links. At least 11 of the 15 answers pointed to one of these sources. In this category ChatGPT behaved less like a friend with an opinion and more like a cautious adviser handing over a shortlist and a phone book.

Home care: franchise networks and the province's own finder

Home care looked much the same. ChatGPT linked 2.9 home care businesses per answer, and franchise networks and national providers took 52% of those links. Home Instead's Canadian site appeared in four of 14 answers, and its US site once. Right at Home appeared in three. Qualicare, Nurse Next Door, Bayshore, FirstLight, C-Care, PSW Direct and Just Like Family appeared in two each.

No category leaned harder on public and directory sources. Of 60 links in the home care answers, 20 went to a finder, a directory or a government site rather than to a company. Ontario Health atHome, the provincial agency that coordinates publicly funded home care, was linked five times. SeniorCarePath was linked three times, and the regional health service finders four. For a private agency, that means a third of the space in these answers goes to sites that send families somewhere else first.

Fertility clinics: the same four names in most answers

Fertility answers were longer, at 3.9 clinics each, but concentrated at the top. Mount Sinai Fertility appeared in seven of 15 answers, TRIO in six, and TWIG and Hannam in five each. Pollin, NewLife and Generation appeared in four. The three most-linked clinics took 31% of all clinic links, about the same concentration as retirement homes.

The government showed up here too. Pages on ontario.ca were linked in seven answers. The province funds some fertility treatment, and ChatGPT appears to treat it as part of the answer to "where should I go", alongside the clinics themselves.

Resorts: Deerhurst, Marriott and Vintage Hotels, then a long tail

Resort questions produced 70 links to 47 different resorts and hotels. Deerhurst Resort in Huntsville, Marriott (whose Ontario properties share one domain, so we counted them together) and Vintage Hotels in Niagara-on-the-Lake each appeared in four of 15 answers. Bayview Wildwood, Blue Mountain, Ste. Anne's and Couples Resort appeared in three. Then the list fans out: 34 of the 47 appeared once, among them Cobble Beach, Arowhon Pines, Fern Resort and Elmhirst's.

The detail resort owners will care about is what was missing. In 15 answers ChatGPT did not link to Expedia, Booking.com or Tripadvisor once. Every link to a place to stay went to the property's own website, and only two answers linked a regional tourism site. In this snapshot, at least, an AI recommendation sent the traveller straight to the resort rather than to an online travel agency that takes a commission on the booking.

Private schools, venues, camps and franchises: everyone gets a turn

The other four industries looked nothing like the care categories. ChatGPT linked five or more businesses per answer and spread them thin. Private schools produced 73 different schools from 15 questions, and 89% appeared once. Bayview Glen was the most-linked school, in three answers. Event venues produced 68 venues, 90% of them once, with Graydon Hall in three answers. Summer camps produced 63, led by Camp Kandalore and the YMCA of Greater Toronto at four each.

Franchises scattered most of all: 72 different brands, 93% named once, and no brand in more than two answers. The most-linked franchise source was not a franchise. The Canadian Franchise Association's site appeared in eight of 15 answers, more than any brand. Directories mattered little for schools: only two answers linked one, the Our Kids guide both times. Venue answers linked a directory in seven of 15, EventSource most often.

In these categories a single question is close to a lottery. Across all 15 venue answers only a handful of venues were named more than once, which is what you would expect when there are hundreds of plausible answers and no regulator to point to.

Why care categories get shorter, more repetitive lists

OpenAI does not publish how ChatGPT chooses businesses, so what follows is our reading of the pattern rather than a finding. Two explanations fit what we saw, and they are not exclusive.

The first is caution. Questions about a parent with dementia or a fertility diagnosis are high stakes, and ChatGPT's answers in those categories hedged more: fewer businesses, more pointers to regulators, provincial agencies and public finders. Naming two homes and the register is a safer answer than naming seven homes.

The second is coverage. The operators that dominated the care answers run many locations and publish a page for each one, so a question about almost any town in the region has a matching page from Amica or Chartwell to find. An independent home has one site about one building. When ChatGPT searches the web to answer a question about Oakville or Pickering, the large operators are simply more likely to have a page that matches.

What a business can do with this

Start by asking the questions yourself. Write down the 10 to 20 things your customers actually ask, with the city or neighbourhood they would name, and put each one to ChatGPT in a fresh logged-out chat. Our questions below can serve as a template. Run them again on another day, because a single answer is noise. If you are named in three runs out of ten, that number means something. Being named once does not.

In the short-list categories, the sources ChatGPT linked were the operators' own websites and the public registers. That makes two things worth checking: that your listing with the regulator or provincial finder is complete and current, and that your own site has a page for each location and service people ask about, in plain words. A question about respite stays in Milton can only be matched to a page that says respite stays in Milton. We cannot prove those pages cause a mention, but they are what was cited.

In the scattered categories the job is coverage across wordings. A venue that is right for 400-guest weddings, small downtown ceremonies and winter events needs to say so, because each of those was a separate question in our set, and the names in the answers barely overlapped.

And measure it over time rather than once. If you want that done on a schedule for your own list of questions, AIO Copilot's AI search view runs it. A single business checking a handful of questions every month can do the same thing by hand for nothing.

How we ran the test, and its limits

We wrote 15 questions for each of eight industries, 120 in all, in the voice of a Greater Toronto Area resident. Most name a city or neighbourhood in the region. Franchise questions are Canada-wide, and resort and camp questions cover Ontario. The industries were asked in rotation, one question from each in turn, so that no industry was asked entirely at one time of day.

Each question went into a new chat in a logged-out ChatGPT session in a browser in Toronto, with no account, history or memory, between about 3:20 pm and 5:40 pm Eastern on 27 October 2026. We waited for each answer to finish and recorded every website domain it linked to, counting each domain once per answer and ignoring map tiles and ChatGPT's own links. We then sorted domains into businesses and into directories, regulators, government sites and review sites. Answers under 400 characters or that failed to load were excluded, which removed one home care answer.

The limits matter more than usual in a study this size:

• Each question was asked once. ChatGPT's answers vary between runs, so a rerun would give different names. Differences of one or two answers between businesses are within that noise.
• We counted links, not mentions. A business named in the text without a link is not counted.
• A domain is not always one business. Marriott's domain covers several hotels, and a few businesses use two domains (Sunrise, Home Instead, TRIO and CReATe among them), which splits their count.
• Logged-out answers can differ from what a signed-in user with saved memory and location sees.
• Which operators count as multi-site, franchise or national is our classification.

All 120 questions we asked ChatGPT

The exact wording, grouped by industry. Use them as a template for your own check.

The leaderboard matters less than the pattern. Where the answer is a short list and a register, the businesses on that list are getting a recommendation their competitors are not, every time someone asks. Where the answer scatters, nobody owns the question yet.

Frequently Asked Questions