How AI Decides Which Local Business to Recommend

In BrightLocal's Local Consumer Review Survey 2026, published on March 10, 2026, 45% of consumers said they had used an AI tool to find a local business in the previous twelve months. The same question a year earlier returned 6%.

That is a survey of 1,002 US adults, with 455 of them in the AI-using subset, so treat it as a strong signal rather than a precise measurement, and note that it is US data. Canadian adoption tends to track a few months behind. The direction is not in doubt regardless. In a year, asking a chatbot for a plumber went from a novelty to something close to half the market having tried it.

That last figure is the one worth pinning to the wall, and we will come back to it.

A recommendation is not a ranking

Ranking has room in it. Position seven on page one still gets clicks, still gets compared, still gets chosen by somebody who scrolls. Ten businesses are on the page and the customer does the sorting.

An assistant asked for the best roofer in Mississauga names two or three. There is no position seven in a sentence. The list has been pre-sorted, and everybody not on it was not rejected by the customer. They were never shown.

This is why local businesses feel the shift harder than most. A 3% share of a ten-link page is a normal outcome. A 0% share of a three-name answer is the same visibility problem with a very different shape.

What Google states on the record

Google publishes guidance on appearing in its generative AI features, and it is unusually direct for this subject. Three things in it matter to a local business.

The AI features run on the same ranking systems as Search. In Google's words, "the best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems". There is no separate AI ranking to game.

Your page has to be eligible for a snippet. A page must meet the ordinary indexing bar and be eligible to appear in Google Search with a snippet before it can appear in an AI feature. If you have nosnippet set anywhere, or your content only assembles in the browser, this is where you fall out, and that is worth checking properly, because it takes a minute and it is binary.

Business Profile is named explicitly. Google's guidance points local businesses at Google Business Profile, and ecommerce at Merchant Center, as the way to make services and products visible "in both AI responses and other Google Search results".

The same document is equally clear about what does not help, and it is worth quoting because a lot of money is currently being spent against it: "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities)." Google names llms.txt specifically and says such files neither harm nor help visibility in Search. We publish one anyway, for assistants that are not Google, and we have never claimed it does anything for Google. Now you have it from the source.

What ChatGPT does, honestly

Nobody outside OpenAI can tell you how ChatGPT picks a local business, because OpenAI does not publish it.

What has been established by testing, rather than asserted, is that ChatGPT answers local questions by running a live search and reading results. Search Engine Land's 2025 side-by-side testing found it pulling roughly the top 20 to 30 web results from Bing and applying its own filtering on top, and found no evidence it reads the profile data businesses manage in Bing Places the way Google reads a Business Profile.

You will also find confident percentages circulating: that some fixed share of its local answers comes from Foursquare, or Yelp, or a maps provider. We went looking for the methodology behind the most-cited of those numbers and there isn't one published. No sample size, no date, no query set. They may well be directionally right. They are not evidence, and building a strategy on an undocumented pipeline that can change in a product update is how agencies end up selling work that quietly stops mattering.

The defensible read is duller and more durable. These systems assemble an answer from whatever the open web says about you: your site, directories, review platforms, local press, social profiles. You cannot optimise the pipe. You can be the business that is easy to describe accurately from any of those sources, which is the same thing from the other end.

One practical consequence: Bing matters more than its search share suggests, because both ChatGPT and Microsoft Copilot lean on its index. If your Bing Places listing is unclaimed or stale, fix it this week. It is twenty minutes and almost nobody in the local market has bothered. Submit your sitemap in Bing Webmaster Tools while you are there. We did ours in September 2026, Bing read all 71 URLs, and it takes IndexNow pings for new pages, which Google does not use.

The thresholds people quote

You will read that you need 4.3 stars and 25 reviews to be eligible for AI recommendation, or similar clean-looking numbers. Be sceptical. Those figures come from small anecdotal test sets rather than published research, and they are stated with a confidence the evidence does not carry.

What is documented is more useful anyway. Reviews are the substrate these systems read to decide whether a business is any good, because it is the only large body of opinion about you that exists in a machine-readable form. A business with fourteen reviews is not banned from being recommended; it is harder to describe confidently, and a system generating a recommendation in one sentence prefers a business it can describe confidently.

So the useful version of the advice is not a number. It is: have enough recent reviews that a stranger reading them learns something specific about what you do, and reply to them, because your replies are also text about your business that gets read.

The verification step is the actual finish line

Come back to that 88%. Nearly nine in ten people who get a recommendation from an AI check it before they act on it.

Which means being named is only half of it. The other half is what happens in the thirty seconds afterwards, when they search your business name and land on your site, your Google Business Profile, and two directory listings at once.

That is where local businesses lose customers they had already won:

• The hours in the AI answer are last year's, because a directory nobody has logged into since 2022 still says so.
• The phone number on the site does not match the one on the profile.
• The business name has three variants across four listings, so the customer cannot tell whether they have found the right company.
• The site was last touched in 2019 and the assistant just described the business as modern and responsive.

Consistency work is unglamorous and it is the highest-leverage hour in this entire discipline. Your name, address, phone number and hours should be character-for-character identical on your site, in your structured data, on your Google Business Profile, on Bing Places, and on every directory that carries you. A model that finds two versions of a fact tends to drop the claim rather than pick one, and a customer who finds two versions assumes the worst.

What to do this month

In order of leverage, for a business that has none of it in place:

1. Complete the Google Business Profile properly. Every field, real categories, real hours, real photos. It is the single largest input into Google's local answers and Google says so itself. The mistakes that cost you are worth reading first, because one popular tactic in that post gets listings suspended.
2. Claim Bing Places. Twenty minutes, and it is the index behind ChatGPT and Copilot.
3. Audit your NAP everywhere it appears. Site, schema, profiles, directories. Fix the mismatches, then stop changing your phone number.
4. Ask for reviews on a schedule, and reply to all of them. Recency counts. A wall of five-year-old reviews describes a business that may not exist any more.
5. Write one page per service area that says something true and specific about that area. Not the same page with the city name swapped, which is a spam pattern and reads like one. This is what our service area pages are for, and they were written that way deliberately.
6. Put your facts in plain sentences on your own site. What you do, where, for whom, since when, at what starting price. An assistant can only repeat a claim it can extract, and marketing voice contains nothing extractable.

None of that is new. That is the point of it. The channel is new, the mechanism underneath is the same local SEO that has worked for a decade, and the businesses being recommended by ChatGPT this year are overwhelmingly the ones that did the ordinary work before anyone was calling it AI.

We do this for businesses across the GTA and remotely across Canada and the US, as part of our AI SEO and analysis service. If you want to know where you stand first, ask an assistant for the best business in your category in your city and see who it names, and then measure what it sends you, because most of it arrives disguised as direct traffic.