How does ChatGPT decide which businesses to recommend?
When someone asks ChatGPT for a plumber, a dentist or a roofer and web search is on, it does not pull a name from memory. It turns the question into a few smaller searches, reads the pages that come back, and builds the answer from them. The businesses it names are the ones those pages talk about.
Ahrefs studied 1.4 million ChatGPT prompts and found that pages retrieved through its search channel were cited 88% of the time, while Reddit threads it read were cited less than 2% of the time. The pages that got cited had titles closer to those smaller searches, and pages with plain language URLs were cited more often (89.8% against 81.1%).
In practice that means two things for a local business. You have to be on the pages it finds, and the pages about you have to say clearly what you do and where, in the words a customer would use.
What did ChatGPT read when we asked it for a plumber?
In October 2026 we asked an OpenAI model through the API, with web search on, the way a customer would: I need a plumber in Austin, TX. Which companies would you recommend? It ran two searches and read ten sources. Most of them were plumbing companies' own websites, plus a consumer review roundup and a city contractor registry.
We asked the same question a second time and the list changed. One company that came up in one run was missing in the other. AI answers vary between runs, so a single check is a snapshot, and the businesses that show up consistently are the ones present on many of the pages it reads, not just one.
You can run the same check for your own business and city with our free AI visibility tool. It shows who gets recommended and which websites the AI read to decide.
What kind of pages get cited instead of yours?
In September 2026 we measured our own company the same way: 13 questions a buyer might ask about services like ours. We were recommended in zero of them. Across the 160 pages the AI assistants cited instead, 36% were articles or guides, 15% were comparisons and only 3% were product pages.
Pages that answer a question get cited. Pages that sell rarely do. A homepage that says you are the best plumber in town is easy to ignore. A page that explains what a water heater replacement involves in your city, and who does it, is something the AI can use.
Step 1: Find the questions your customers actually ask
Write down the ten questions people ask you on the phone before they book. Then type the start of each one into Google and look at the suggestions it offers: those are real searches, in the words people use.
Each of those questions is a candidate page or a candidate section on your website. The title should be the question, and the first paragraph should answer it.
Step 2: Be on the pages the AI reads for your trade
Run the check for your trade and city and look at the sources list. If it reads a review site, a directory or a local best of list, that is where you need a complete, accurate listing.
Your Google Business Profile matters here too: categories, services, hours, service area, photos and answered questions.
Step 3: Make your details identical everywhere
Same business name, same phone number, same address or service area, same list of services on your website, your Google profile and every directory. When sources disagree, the AI has less reason to trust any of them.
Step 4: Collect reviews every week, from every customer
Review sites and review roundups often show up among the pages the AI reads, and reviews give it the specifics it uses as reasons: fast, fair price, cleaned up after. Ask every finished customer with a one tap link, right after the job.
Ask everyone, including the customers you are less sure about. Google's policy prohibits selectively asking for positive reviews.
Step 5: Give each page a title that is the question
Use the customer's own words in titles and URLs: emergency-plumber-austin, not services-page-3. Put the short answer in the first paragraph, then the details. This is the same structure that helps a page rank in normal search.
Step 6: Check every month and track the visits
Run the same question once a month and keep a note of who is named and which sources were read. Changes in the sources list tell you where to work next.
Links that ChatGPT shows carry a tag your analytics can read (utm_source=chatgpt.com), so you can see how many visitors arrive from it. Ask new customers how they found you as well; some will say they asked an AI first.
What changes little on its own
Adding schema markup alone, or paying for a report that only scores your website, rarely changes who gets recommended. They do not put you on the pages the AI reads. We publish an llms.txt file on our own site, but we have not seen evidence that it changes recommendations by itself.