AI visibility for local businesses: get named by ChatGPT

Thursday, 10:05 p.m.: one question, three names

Thursday, 10:05 p.m. A tenant in Sherbrooke watches water back up through the basement floor drain. She asks ChatGPT: "Which plumber in Sherbrooke can come Sunday morning for a blocked drain?" The answer names three businesses, one sentence each. Yours isn't one of them, even though you've done exactly that, on Sundays, for twelve years.

It isn't necessarily that the AI thinks you're worse. More often, it simply couldn't find enough sources that agree about you. Microsoft says so plainly when describing how Bing feeds AI answers: abstaining is a valid outcome when information is missing, stale or conflicting (Bing, 2026). An AI assistant doesn't recommend the best business. It recommends the one it can describe without getting it wrong.

That's the point of this article. You can't "optimize for ChatGPT," but you can make yourself easy to describe. In practice, the six facts (name, trade, phone, hours, services, languages served) should match everywhere you're mentioned, in both of your customers' languages. And you can measure where you stand in one evening.

Where AI assistants get what they say about you

The largest public study on this comes from BrightLocal: more than 200,000 local searches and nearly 2 million citations across ChatGPT, Google's AI Mode and AI Overviews (BrightLocal, 2026). Business websites supplied 42% of citations, Google Business Profile 28.6%, Yelp 9.5% and Facebook 2.2%. Your website and your Google profile carry most of the load.

Yelp deserves its own line, even if few of your Quebec customers ever open it. In July 2026, Yelp announced a licensing deal so ChatGPT can show its reviews, ratings, photos and business details (Axios, 2026). Perplexity started adding Yelp data to its local answers back in 2024 (Maginative, 2024). A forgotten Yelp page with an old number is a source contradicting you in exactly the wrong place.

Beyond that, OpenAI refers to "third-party providers" of nearby business listings without naming them (OpenAI), so be wary of guides that claim to know where ChatGPT gets its listings. Two things are documented: your site has to let in OpenAI's search crawler (OAI-SearchBot) and Perplexity's (PerplexityBot) (OpenAI, Perplexity), and Bing Places, which is free, can import your Google listing (Bing, 2025).

The 5-3-2 test: thirty answers, one scorecard

Before you change anything, measure. The 5-3-2 test takes a little over an hour: 5 questions, 3 assistants, 2 languages, so 30 answers.

1. Write 5 real customer questions, the way people ask them on the phone: a service, a town or neighbourhood, a constraint ("Sunday," "fast," "speaks French"). Never your business name: you're testing whether the AI thinks of you, not whether it knows you exist.
2. Pick 3 assistants: ChatGPT, Perplexity and Google's AI Mode.
3. Ask each question in English, then in French, in a private window, signed out. ChatGPT can use saved memories to rewrite a search and uses an approximate location based on your IP address (OpenAI), so always name the town.
4. Score every answer in a plain spreadsheet, using this scale.

Score	What the answer says	What it means
0	You're not named	The AI doesn't link you to this request
1	Named, vague description	It knows you, without enough detail
2	Named, accurate description	Your sources agree
E	Named with an error (hours, number, service, town)	A source contradicts you

Then work out your presence rate: answers that name you ÷ 30. For context, a business tracked by BrightLocal showed up in 32% to 38% of AI searches on average, against 66% on Google Maps (BrightLocal, 2026). Here are my thresholds, in order:

• Any E at all: that's your priority. Find the source that's wrong and fix it before anything else.
• Fewer than 5 out of 30 (under 15%): AI assistants barely know you. Foundations first: website, Google profile, Bing Places.
• 5 to 10 out of 30: the working zone, below the study's average. Work on the six facts and on your weaker language.
• More than 10 out of 30 (over a third): at or above the study's average. Rerun the test monthly and protect what you have.

Add the gap rule, made for Quebec: if one language gets fewer than half the mentions of the other, that's your next project. An English-speaking customer in Lennoxville asks in English; if nothing about you exists in English, you give the answer she gets very little to work with.

Example: a Sherbrooke plumber, before and after

Take a fictional three-person plumbing business in Sherbrooke. Its five questions: a weekend emergency, a water heater replacement, a drain unclogging in Fleurimont, a bathroom renovation in Magog and, for customers in Lennoxville, a plumber who serves people in English.

First test: it's named in 6 of 15 French answers (40%) and 1 of 15 English answers (7%). Presence rate: 7 ÷ 30 = 23%, in the working zone. But two of the seven mentions get an E: one says it's closed on weekends, another gives an old number. And the gap rule flashes red: 1 mention against 6, well under half (3).

The diagnosis takes twenty minutes. The Facebook page still shows the hours from before Sunday service started. An old directory listing has the number from four years ago. The website is French only, and none of the reviews are in English. The plan: fix Facebook and the directory, add an "Emergency plumber in Sherbrooke and Lennoxville" page, import the Google listing into Bing Places, and ask three recent English-speaking customers for a review.

Six weeks later, same questions, same method. Say it gets 7 of 15 in French and 4 of 15 in English: 11 ÷ 30 = 37%, no errors, and the gap closes to 4 against 7, above half (3.5). These numbers are an illustration, not a promise. The real win is the two Es that disappeared: an AI that cites you with the wrong number costs you more than one that doesn't mention you at all.

What most advice gets wrong

The magic file. Someone will sell you an llms.txt file or "AI markup." Google says there are no additional requirements to appear in its AI features, and no new AI text files or special markup to create (Google Search Central). Google's Gary Illyes has also said Google doesn't support llms.txt (Search Engine Land, 2025). That's not where a plumber's or a clinic's visibility is decided.

The ranking. "You're number 3 in ChatGPT" means close to nothing. When SparkToro asked the same questions dozens of times, ChatGPT and Google's AI returned the same list of brands less than once in 100 runs, and the same order about once in 1,000 (SparkToro, 2026). What you can measure is how often you appear. Hence 30 answers instead of one screenshot.

The invisible block. A misconfigured firewall or content delivery network can turn AI crawlers away without anyone noticing. OpenAI notes that your host or CDN must allow traffic from its search crawler's IP addresses (OpenAI). Have whoever runs your website check.

Neglecting the Google profile. Chasing ChatGPT while your Google Business Profile gathers dust means ignoring the second most-cited source in BrightLocal's study. Google itself asks you to keep your Business Profile up to date for its AI features (Google Search Central). If yours is thin, start with the seven Google Business Profile settings.

What you'll never control

OpenAI puts it in writing: ChatGPT ranks results on several factors, and placement is not guaranteed (OpenAI). No package buys a spot in an answer; walk away from anyone who promises one.

Answers also shift from one run to the next. In BrightLocal's study, two identical ChatGPT searches shared only 23% of the businesses they named, on average, and a business that appeared once came back only about half the time (BrightLocal, 2026). ChatGPT also cast a wide net: 90% of its picks fell within 287 km, compared with 39 to 64 km for Google's tools. Going from 23% to 30% in a month may just be noise, so judge on two or three tests.

Finally, that study didn't cover Perplexity, and no study replaces your own numbers. If your work comes mostly from referrals or public tenders, this test can wait. And anything involving your firewall or your site's code belongs with your web developer.

Your test night next week

1. 10 minutes: write your 5 questions, based on the last calls you got.
2. 35 minutes: ask all 3 assistants, in English then French, and score the 30 answers.
3. 10 minutes: work out your presence rate, count the Es, apply the gap rule.
4. 15 minutes: fix the first wrong source, then compare the six facts on your website, Google profile, Bing Places, Facebook and Yelp.
5. In a month: rerun the test with the same questions and compare.

If your site is set up in Bing Webmaster Tools, look at the AI Performance report, still in preview: it counts how often your pages are cited in Copilot and Bing's AI summaries (Bing, 2026).

And if you only do one thing, hunt for the E. A wrong answer about your business is the one AI problem you can definitely fix this week.