An experiment that shows what ChatGPT names for "best [trade] in [city]" prompts across 88,186 businesses (and what it says about getting your business named when buyers ask AI).
Every local business we work with at Complete SEO asks the same question in some form. When a homeowner types "best plumber in Austin" into ChatGPT, who gets named, and why them? Most answers to that are a guess, so in the first week of September 2026 we asked ChatGPT "best [trade] in [city]" 44,988 times, across 35 home service trades and 24 metros, and kept every answer. We ran a second study of 12,120 answers over three days in August to measure how much an answer moves on its own, and we captured 60 answers by hand from a logged-in account to check the instrument. Here's what came back.
95% of the 27,405 answers in the main analysis carried a places card, a short list of businesses with addresses, phone numbers, star ratings and review counts, pinned to a map. The card's business identity tracks the Google listing, and the rating next to a business is Yelp's for the plurality, about half of businesses listed on both platforms against a third for Google. Across all 44,988 answers ChatGPT named 88,186 different businesses, and 85% of them operate in a single market. Ask the same question twice, minutes apart, and the same list comes back 3.6% of the time.
How we asked
The prompts took seven shapes, from "best [trade] in [city]" to a paragraph that asks for a locally owned company and rules out chains, franchises and directory sites. The 35 trades are the home service categories we serve or sell into, roofing to pool care, and they weren't sampled from any wider population, so nothing here is a claim about local services in general. The 24 metros were drawn at random from a seeded sample of US metropolitan areas. Every prompt ran five times, once a day from September 3 to September 7, through DataForSEO's ChatGPT scraper with the location set to the United States and the city named in the prompt, and we didn't simulate a user standing in that city. The main analysis is the 27,405 answers for home service trades in metros with at least five providers of that trade, and the rest sit in a secondary arm that's never merged in.
Business names come from our own extractor, which reads the card and the prose. Against 150 answers coded by hand it found 92.6% of the businesses at 94.1% precision, while the vendor's own entity field found 43.1% of them, so we didn't count from that field. A business is one name in one metro, which gives 96,146 names across the corpus, and 88,186 once the same brand in several metros is folded to one.
Nearly every answer carries a card
99.7% of the 27,405 answers named at least one specific business, 0.2% named only platforms or directories, and none declined and offered selection criteria in place of names. 95% carried the places card, and by trade the range was 92% to 98%. Across the 276,628 businesses those cards held, the median business had 4.9 stars on 28 reviews. On a separate arm where we asked the scraper to turn web search off, the card still appeared on 95.5% of answers, so it isn't a product of a web search. Its rows carry a name, an address, a phone number and a category, which is the shape of a listing.
Where that listing comes from is only half observable. The card's business identity tracks Google's listing data, and the payload doesn't say which provider supplies that layer. The rating is a different story, because each row carries a provider mark. Among card businesses listed on both Yelp and Google, the star rating and review count next to the business matched Yelp's for 48% to 54% and Google's for 26% to 41%, depending on how much drift the matching rule allows between our collection and the comparison pull. Under every rule Yelp leads without reaching a majority, and the split moves with the size of the business. On a panel of 479 card businesses, those with five or fewer reviews carried Yelp's number 71% of the time and Google's 15%, and above 100 reviews it was 22% Yelp and 30% Google.
What a user sees is narrower than what the scraper returns. On a logged-in account the visible card shows about five businesses before the "Expand" control, and the full pool behind it runs to about ten. The scraper returns the pool, so every number in this study describes the pool. On the visible five, Yelp's mark was on 27.9% of rows across 60 captures, and the Yelp-rated businesses sat disproportionately behind the fold.
Who gets named
The cards across the 27,405 answers in the main analysis held about 27,000 distinct businesses, counted by website domain where the row carried one and by name within the metro where it didn't. Add the businesses named only in the prose and the corpus-wide count rises to 88,186. Of a coded sample of 238 of those businesses, 85% operate under their name in one metro only (95% CI 81% to 90%), and 15% operate under the same name in more than one. On the cards, 12% of businesses appeared in more than one of the 24 metros' cards, 62% had an address in the metro's principal city and 95% inside the metro, 56% carried the trade asked about as their category, and 64% carried a website.
The pool is local, and it's deep. Across the seven phrasings an answer named between 3.4 and 10.6 businesses, and in the three-day study one prompt, run more than 200 times, surfaced about 28 different businesses that each appeared in at least 5% of runs. The median one appeared in 30% of runs, and only 36 of 1,115 business-and-prompt pairs appeared in more than 95% of runs, so there's no number one to be, only a pool to be in.
The second road runs through the prose
The card is the main road into an answer, and the prose is the second one. 70.4% of the answers that carried a card also named at least one business the card didn't contain. About three in four of the businesses an answer named were on its own card, and 62% of them sat at card rank five or better, which is roughly the visible part. The rest arrived through the prose, and the prose is written from web pages.
The source domains attached to the 41,160 main-arm answers were led by angi.com (16.5% of answers), expertise.com (12.8%), bbb.org (12.3%), reddit.com (11.3%), homeadvisor.com (8.7%), birdeye (7.7%) and houzz.com (5.4%). On the 595 answers where the payload showed both what was retrieved and what was cited, directory domains were 40% of what came back from the search and 30% of what the answer cited, so a business's own site converts retrieval into a citation better than a directory page does. The searches behind the answer have a shape as well. On a logged-in account 95% of captures carried the engine's own search strings, and they read like "best med spas Phoenix AZ reviews 2026". On the 877 scraper records that kept them, the city was present in 91.7%, the trade in 96.7%, and the word "review" in 31.9%.
How you ask changes who is named
Changing nothing but the wording moved the answer. "List three" returned 3.4 businesses per answer, the plain question 5.8, and the paragraph that asks for a locally owned company and rules out chains and directories 10.6, with the share of answers citing a directory falling from 94.9% to 46.1%. Under that phrasing 98% of answers named at least one business we could verify exists and that met the exclusion, on a random sample of 600. In the three-day study, the long prompts that describe a job and ask whether a quote is fair got a card 78% to 85% of the time against 90% to 98% for the short ones. One phrasing measures one door, and the phrasing results get their own post.
Ask twice, get a different list
Two runs of the same prompt minutes apart, 188,520 pairs across the three-day study, returned the identical set of named businesses 3.6% of the time and the same business first 45.7% of the time. Over three midweek days the same untouched prompts moved by 28 points on their own, before anyone changed anything. Our own standard had been five runs per prompt, and our own measurement says five runs isn't enough to call a change real. Being named in ChatGPT is a rate you measure across runs and phrasings, and any single check is noise.
What this says about getting named
The study measured what a local answer is made of. It didn't measure what moves a business onto the card or up it, and a ranking score sits on every card row that we haven't modeled yet. What follows is our reading of the measurements, labeled as such.
The card holds the listing. It carries listing-shaped identity and fires with web search off, and that much is measured. Our reading is that a listing with an address in the named city and the trade as its primary category meets the entry conditions we saw on 62% and 56% of card rows.
Your Yelp profile is in the answer whether or not you work it. Yelp's rating and count are the plurality on the card pool, and that's measured. Our reading is that a Yelp profile with a strong rating and count is visible in ChatGPT today, and that Google's number takes over as the review footprint grows.
The prose road runs through directories and your own site. The source domains above are measured, and so is the finding that business sites convert retrieval into citation better than directory pages do. Our reading, which this study didn't test, is that presence on those lists is the second road, and whether being listed on any one of them raises your odds isn't something we measured.
A readable site is the cheapest fix on the list. In a separate census of 15,834 live independent local business sites, 23.8% couldn't be read by AI crawlers, and about 9% of all sites refused them outright through robots rules or blocking at the edge.
Measure how often you're named, with a band around it. The 3.6% identical lists and the 28 points of drift in three days are measured. Our reading is to report an appearance rate across runs and phrasings and to treat any single check as noise.
This is the second study in a series from our AI visibility tracking work, after the one that measured where Claude's web search gets its sources, and the next two take the phrasing results and the card itself apart in turn. If you want to know how often your business is named when buyers ask AI, talk to us.
