How we measure whether AI recommends you
Everything on our site and in our reports is a count over logged AI answers, not a screenshot or a guess. Here's exactly what we do, what the numbers mean, and what they can't tell you.
What we ask, and who we ask
For each industry we build a set of questions a real customer asks before hiring, from the first "is it worth repairing" through "what does it cost" to "who should I call." For heating and cooling in the GTA that's 16 questions, chosen from search-volume data and published research, phrased the way a Toronto homeowner would type them.
We put every question to four AI assistants: ChatGPT, Google's Gemini, Claude, and Perplexity. Each question is asked three times per assistant, because the same question gets different answers on different runs. Every answer is saved word for word, with the date and the list of web pages the assistant said it used.
16 questions, 4 assistants, 3 runs each is 189 logged answers for the September 10, 2026 study (three runs failed and were excluded rather than guessed).
What the numbers mean
| Word we use | What we counted |
|---|---|
| Named | Your business name appears in the answer text. |
| Cited | Your own website is among the sources the assistant says it used. |
| Appeared | Either of the above. This is the headline number in our reports. |
| Recommended | The answer tells the customer to pick you. This one is a judgement, so it's classified from the logged text and every classification keeps the exact quote. |
| Position | Where you sit among the other tracked businesses named in the same answer. First is first. |
| Source gap | A page the assistants read often that names your competitors and not you. |
A business's share is its appearances divided by the sampled answers. Every business in an industry is measured against the same logged answers, so shares are directly comparable.
What it can't tell you
- It's a sample, not a census. 189 answers on 16 questions. Other phrasings would give other results. Movement of fewer than 2 appearances on a question is noise, and we'll always call it noise.
- Assistants change. Published testing found AI engines agree on their top recommendation for the same question only about 44% of the time. That's why we never report a single run.
- Toronto-phrased questions favour Toronto businesses. A zero for a Mississauga-only firm partly reflects geography. Cost and emergency questions are less affected.
- AI is one channel among several. It sends a small share of most businesses' traffic today. This sits alongside Google and referrals; it doesn't replace them.
The numbers on our home page, and where they come from
| Number | Source |
|---|---|
| 41% of Ontarians (39% of Canadians) used AI in the past year to support a purchasing decision | Léger survey for National Bank, June 5 to 7, 2026, 1,518 Canadian adults |
| 45% of people used AI to find a local business last year, up from 6% the year before | BrightLocal Local Consumer Review Survey, February 2026, 1,002 US adults |
| About 1 in 80 businesses is recommended by ChatGPT vs about 1 in 3 appearing in Google's local map results | SOCi Local Visibility Index 2026 (1.2% vs 35.9%, 350,000+ locations), reported by Search Engine Land, January 2026 |
| Roughly two thirds of the sources AI relies on are third-party pages, not the business's own site | Omniscient Digital citation analysis (23,000+ citations) and Whitespark local AI citation sample, 2025 to 2026, which put the range at 60 to 77% |
| 23 of 48 GTA heating and cooling companies were never named in 189 answers; two companies took 36% of all mentions | Our own study, probed September 10, 2026, by the method above |
| AI engines agree on the top recommendation only about 44% of the time | 37,000-run cross-engine consistency study across 533 brands, 2026 |
How the free check works
Tell us your business name and website. We run the same questions for your industry (or read your line from a study we've already run), check whether your website is blocking the AI crawlers by fetching it the way each crawler does, and look at whether you're on the pages the assistants relied on. You get one email within a business day with what we found and a few fixes you can do yourself. We keep the raw answers so any number in that email can be checked.
Questions about the method? hello@getpotentia.ca.