The local businesses AI recommends have better-built websites
We checked 1,056 local business websites in six Canadian cities, then asked AI assistants for recommendations. The businesses they named had better-built sites.
In October 2026 we checked a random sample of 1,056 local business websites in Toronto, Montréal, Vancouver, Calgary, Ottawa and Edmonton, across 11 kinds of business from dentists to roofers. We then asked ChatGPT, Gemini, Google AI Mode and Perplexity to recommend local businesses, and checked the websites they pointed to in the same way.
Key findings
- AI assistants look for local businesses in different places. Gemini and Google AI Mode answered mostly from Google Maps (77% and 70% of their answers cited it). ChatGPT used the widest mix: businesses' own websites in 80% of its answers, and Google Maps, "best of" lists and review sites in just under half each. Perplexity cited businesses' own websites and the Better Business Bureau most.
- The businesses AI recommended had better-built websites. Their median score was 72 out of 100, against 64 for a random sample of the same kinds of businesses. Only 22% left out business details for search engines, against 47% of the random sample.
- Most local business websites miss at least one basic. 61% had pages without a description for search results, 51% had pages without a main heading and 47% didn't describe the business in a form search engines and AI read.
- Hardly anyone blocks AI. Only 2 of 716 sites blocked a crawler that AI assistants use while answering, and 8 blocked a crawler that collects AI training data.
- About 1 in 7 listed websites no longer exists. For 162 of the 1,056 businesses (15%), the website in public business listings pointed to a domain that no longer works.
What AI assistants cite when you ask for a local business
We asked each assistant the same question for 11 kinds of business in each city, such as Can you recommend a good plumber in Calgary, Alberta?, and the 11 questions again in French for Montréal. Each question was asked twice, which gave 614 answers on 11 October 2026. For each answer we recorded the pages it linked to as sources.
| Assistant | Google Maps | Review sites and directories | "Best of" lists | A local business's own site | Reddit and forums |
|---|---|---|---|---|---|
| Gemini | 77% | 7% | 13% | 18% | 1% |
| Google AI Mode | 70% | 15% | 5% | 16% | 11% |
| ChatGPT | 47% | 46% | 47% | 80% | 2% |
| Perplexity (API) | 0% | 34% | 10% | 48% | 0% |
The review sites, lists and forums cited most often were HomeStars, the Better Business Bureau, TrustAnalytica, ThreeBestRated, Reddit, Yelp, BestProsInTown, Houzz and Birdeye. Which ones matter depends on the trade: HomeStars came up most for roofers, electricians and renovation contractors, Houzz almost only for renovation, and the Better Business Bureau most for roofers and heating companies.
Asked in French in Montréal, the assistants cited "best of" lists less often than in English (7% of answers against 20%).
The businesses AI recommends have the basics in place
We then checked the websites the AI answers pointed to, with the same checks as the random sample. We kept only businesses from the same group the sample was drawn from: independent businesses in the 11 categories and six cities, recommended for their own category. That gave 244 websites; 231 could be fully checked.
Show as a table
| Failing | Random sample | Recommended by AI |
|---|---|---|
| No business details markup (Organization or LocalBusiness) | 47% | 22% |
| No service details markup (Service, Product or similar) | 85% | 67% |
| Pages without a description | 61% | 45% |
| No link preview tags | 68% | 53% |
| Pages without a main heading (H1) | 51% | 37% |
| Missing or repeated page titles | 34% | 21% |
| No redirect to HTTPS | 13% | 2% |
| No sitemap | 14% | 5% |
| No language or mobile viewport tag | 10% | 5% |
Overall, the recommended businesses' websites had a median score of 72 out of 100 on Connu's 25 checks, against 64 for the random sample, and 25% of them scored 80 or more, against 13%.
What this shows, and what it doesn't
This is a correlation, not proof that fixing a website gets a business recommended. The businesses AI recommended also look busier: 32% of their sites had enough visitors for Google to publish real-visitor speed data, against 11% of the random sample. Busier, more established businesses tend to invest more in their websites, and they also tend to have more reviews and listings, which AI assistants cite too.
Most local business websites miss at least one basic
In the random sample, 716 of the 1,056 websites could be fully checked. These are the shares that failed each check:
| Check | Sites failing | Why it matters |
|---|---|---|
| Service details markup | 85% | Tells search engines and AI what services the business offers. |
| Link preview tags | 68% | Controls the title and image people see when the site is shared. |
| Page descriptions | 61% | The short summary search results can show under the title. |
| Main heading (H1) | 51% | Says what each page is about. |
| Business details markup | 47% | Name, address, phone and type of business in a form machines read. |
| Page titles, missing or repeated | 34% | The clickable headline in search results. |
| Image descriptions (alt text) | 29% | Lets search engines, AI and screen readers know what images show. |
| Sitemap | 14% | Lists the pages for search engines to find. |
| HTTPS redirect | 13% | Sends visitors to the secure address instead of a "Not secure" page. |
| Language and mobile viewport | 10% | Tell browsers the page's language and how to fit it to a phone screen. |
By city, no city stood out: the median score was 64 in five cities and 68 in Ottawa.
By trade, page descriptions were missing most often on physio and chiro clinic sites (79%) and renovation contractors' sites (76%), and least often on roofers' sites (41%). Heating and cooling companies were the most likely to describe their business in markup (28% missing, against 58% for auto repair shops and renovation contractors). Each trade had 57 to 75 fully checked sites, so a gap of less than about 17 points between two trades could be chance.
By website platform, the platform's defaults show. 37 of 38 Squarespace sites had pages without a description, and 24 of 34 GoDaddy sites didn't describe the business in markup. Only 1 of 72 Wix sites failed to redirect to HTTPS. WordPress ran 62% of the sites whose platform we could tell (328 of 525).
Speed on phones: the lab test and real visitors
Google's PageSpeed test loads a page the way a mid-range phone on a slow 4G connection would. On that test, the main content of the median homepage took 7.6 seconds to appear, and 95% of the 770 homepages measured were slower than Google's 2.5-second mark.
The lab test is stricter than most visitors' experience. Google also publishes speed data from real Chrome users, but only for sites with enough visitors: 11% of the sample. Among those, about half (52%) met the 2.5-second mark for three out of four visits. Among the AI-recommended businesses with this data, 69% did, but they were no faster on the lab test, so the gap is more likely about who visits these sites than how the sites are built.
Hardly anyone blocks AI crawlers
Websites can tell crawlers to stay away in a file called robots.txt. Of the 716 sites, only 2 blocked a crawler that an AI assistant uses while answering (both blocked ChatGPT-User, which ChatGPT uses to open a page). None blocked OpenAI's OAI-SearchBot, PerplexityBot, Claude-SearchBot or Bing's crawler. 8 blocked at least one crawler that collects AI training data, most often Common Crawl's CCBot and OpenAI's GPTBot. When a local business is missing from AI answers, blocking is rarely the reason.
About 1 in 7 listed websites no longer exists
Our sample came from public business listings (the Overture Maps open places data). For 162 of the 1,056 businesses (15%), the listed website's domain no longer resolved, and a second check from another network gave the same result. Listings like these feed map apps and directories, so a dead link there can send customers, and AI assistants, nowhere.
What local businesses can do
- Keep your Google Business Profile complete, with the right website address, hours and categories. Gemini and Google AI Mode answered mostly from Google Maps.
- Be listed and reviewed where AI looks in your trade, such as HomeStars, the Better Business Bureau, Yelp or Houzz. Ask an assistant the question your customers ask and see what it cites.
- Fix the basics on your own site: a title and description on every page, your business details in structured data, a redirect to HTTPS and a sitemap.
- Don't block AI search crawlers in robots.txt unless you mean to.
- Check the website address in your listings, and fix any that point to an old domain.
You can check your own website against the same 25 checks with Connu's free website score.
How we did the study
- Sample. We took every place with a website in the Overture Maps open places data for Toronto, Montréal, Vancouver, Calgary, Ottawa and Edmonton (downloaded on 10 October 2026), in 11 categories: dentists, physio and chiro clinics, law firms, accountants, plumbers, electricians, heating and cooling, roofers, renovation contractors, auto repair, and hair salons and barbers. We removed chains (a brand, or three or more locations sharing a site), social media and directory pages, and duplicates, which left 21,546 independent business websites. We drew 16 at random from each city and category, 1,056 in total.
- Website checks. On 11 October 2026 we ran Connu's 25 website checks on each site: a crawl of up to 20 pages that follows robots.txt, as ConnuBot, plus Google PageSpeed Insights on the homepage (median of up to three runs per device). 716 sites were fully checked, 120 only for speed (our crawler couldn't read their pages), and 220 couldn't be reached, 162 of them because the domain no longer exists.
- AI answers. 77 questions (66 in English, 11 in French for Montréal), each asked twice of ChatGPT, Gemini and Google AI Mode, captured from the apps as people use them with Canada as the location (through DataForSEO), and of Perplexity through its API with web search, answering with the openai/gpt-6-luna model and no set location. Every question names the city. That makes 616 asks; 2 failed. We classified the cited pages by their address, and counted a business website only when it matched a business in the listings.
- Comparison. The recommended group is the 244 cited business websites in the same listings, trades and cities, recommended for their own trade (26 of them were also in the random sample). Differences are reported only where a 95% bootstrap interval excludes zero.
- Limits. The listings don't include every business. AI answers vary from one ask to the next, and two asks per question is a small sample. Perplexity's API may answer differently from its app. Outside Montréal we asked in English only. The comparison shows a correlation, not a cause.
Business owners who would like their site removed from the study can write to info@connu.ai. We publish only totals and never name a business in the study.
Use the data
399 rows, totals only: no business is named. Each row is one figure: a share (count of total), a median, or the difference between the two groups with its 95% interval.
The figures and charts are free to use under CC BY 4.0, with credit to Connu and a link to this page.
To cite: Connu, “The local businesses AI recommends have better-built websites”, October 11, 2026, https://connu.ai/research/canada-local-business-websites-2026