We studied which businesses ChatGPT names when asked "best [trade] in [city]" across 88,186 businesses, and what the findings mean for owners who want their own business to come up when a buyer asks an AI tool.
Every local business client we work with asks us which businesses ChatGPT names when a homeowner types "best plumber in Austin," and why it names them. Most answers out there are guesses, 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 built a separate study of 12,120 answers across three days in August to measure how answers shift when nothing changes, and pulled 60 answers from a logged-in ChatGPT account to check our measuring instrument.
In the main-analysis subset of 27,405 answers, 95% included a places card, a short business list with addresses, phone numbers, star ratings, and review counts pinned on a map. The card follows a business's Google listing for its identity, and among card businesses listed on both Yelp and Google, about half showed Yelp's rating and about a third showed Google's. Across all 44,988 answers, ChatGPT named 88,186 distinct businesses, and we estimate from a sample of 238 of those businesses that 85% operate in one market only. In pairs of runs from the three-day study, the same question asked minutes apart returned an identical list only 3.6% of the time.
How we asked
Our seven prompt shapes covered 35 home service categories we serve or sell into, from roofing to pool care, ranging from "best [trade] in [city]" to a paragraph requesting a locally owned company and excluding chains, franchises, and directory sites, though those trades weren't sampled from a broader population and the findings don't speak for local services as a whole. We randomly drew 24 metros from a seeded sample of US metro areas. Each prompt ran five times, once per day from September 3 through September 7, and we collected answers using DataForSEO's ChatGPT scraper with the location set to the United States and the city only in the prompt text, without simulating a user physically in that city. The main analysis includes 27,405 answers for home service trades in metros with at least five providers of that trade, while all other answers went to a secondary arm and were never merged into the main analysis.
Our own extractor pulled business names from both the card shown in an answer and its written response, finding 92.6% of the businesses with 94.1% precision on a check set of 150 hand-coded answers. On the same 150 answers, the vendor's entity field found 43.1% of businesses, so none of our counts come from that field. We counted one business as one name in one metro, yielding 96,146 names across all the collected answers. Folding the same brand across several metros into one brings the count to 88,186.
Nearly every answer carries a card
Of the 27,405 answers in our main analysis, 99.7% named a specific business, 0.2% named only platforms or directories, and none declined to name businesses while offering selection criteria. A places card appeared in 95% of answers, with rates by trade from 92% to 98%. The cards held 276,628 listings in all, and the median business had 4.9 stars and 28 reviews. Each card row listed a name, address, phone number, and category, the shape of a business listing. In a separate arm, the scraper was asked to switch web search off, yet a card appeared in 95.5% of answers, so the card doesn't come from a web search.
![A ChatGPT answer to “best [trade] in [city]” with its places card of rated map pins and business rows, and a note that 95% of answers in the main analysis carried a places card.](/images/blog/inline/chatgpt-local-answers-study-places-card.png)
Only the rating source was visible to us through a mark on every row, while the card's business identity followed Google's listing data without a named provider in the payload. Among card businesses with profiles on both Yelp and Google, the card's star rating and review count matched Yelp's in 48% to 54% of cases and Google's in 26% to 41%, depending on how much drift the matching rule allowed between our collection and the later comparison pull. Yelp came out ahead under every matching rule, and the split shifted with review count. In a rating panel of 479 card businesses, those with five or fewer reviews showed Yelp's number 71% of the time and Google's 15%, while above 100 reviews, the figures were 22% for Yelp and 30% for Google.
On a logged-in account, the visible card showed a few businesses before the "Expand" control, with about ten in the full pool behind it. The scraper returned that full pool, which every figure in our study describes. Across 60 logged-in captures, Yelp's mark appeared on 27.9% of visible rows. Businesses with Yelp ratings sat disproportionately below the fold.
Who gets named
Cards in 27,405 main-analysis answers named roughly 27,000 distinct businesses, counted by website domain when present or by name within the metro otherwise. Adding prose-only mentions raised the corpus-wide count to 88,186. In a sample of 238 businesses from that group, 85% traded under their name in only one metro, with an estimated range of 81% to 90% at 95% confidence, and 15% traded under the same name in two or more metros. Among card businesses, 12% appeared in more than one of the 24 metros, 62% had an address in the principal city, and 95% inside the metro, while 56% listed the requested trade as their category and 64% listed a website.
The pool is local and deep. Across the seven phrasings, averages ranged from 3.4 to 10.6 businesses named per answer. In the three-day study that ran one prompt over 200 times, about 28 different businesses each appeared in at least 5% of runs, and the median among them appeared in 30%. Only 36 of 1,115 business-and-prompt pairs appeared in 95% or more of runs, so there's no number-one spot to win and the aim is to be in the pool.
The second road runs through the prose
The card is the main road into an answer, but businesses can also enter through the prose ChatGPT writes from web pages. Among answers with a card, 70.4% named at least one business that wasn't on it. About three in four businesses named in an answer appeared on that answer's own card. The rest came through the prose.
For the 41,160 main-arm answers, a separate base from the 27,405 used in the main analysis, source domains attached to answers included 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%). In 595 answers that showed both retrieved and cited results, directory domains accounted for 40% of retrieved search results and 30% of cited ones, and a business's own website converts retrieval into a citation at a better rate than a directory page. All 60 captures from a logged-in account showed the engine's search strings, with "best med spas Phoenix AZ reviews 2026" as one example. Among 877 scraper records that kept those strings, 91.7% included a city, 96.7% included a trade, and 31.9% included the word review.
How you ask changes who is named
With only the wording changed, "List three" produced 3.4 businesses per answer, a plain question produced 5.8, and a long paragraph asking for a locally owned company while excluding chains and directories produced 10.6. With the move to the long, locally owned phrasing, the share of answers citing a directory fell from 94.9% to 46.1%. In a random sample of 600 answers to that phrasing, every answer named at least one business we could confirm existed and met the exclusion. In the three-day study, cards appeared in 78% to 85% of answers to long prompts describing a job and naming a particular service, compared with 90% to 98% for short prompts. Each phrasing shows one way into the answer, and we'll cover the phrasing results in a separate post.
Ask twice, get a different list
Across 188,520 pairs of runs of the same prompt minutes apart in the three-day study, the runs named an identical set of businesses in 3.6% of pairs and the same business first in 45.7% of pairs. Across three midweek days with no change made, two scans of identical prompts had to differ by about 28 points before the difference exceeded the prompts' own drift. We used to run each prompt five times, but our measurement now shows that five runs can't establish whether a change is real. Being named in ChatGPT is a rate measured across many runs and phrasings because any one check is noise.
What this says about getting named
The study measured the makeup of a local answer. It didn't measure what puts a business on the card or raises its position there, and we haven't modeled the ranking score carried by every card row yet. The practical takeaways that follow are our interpretation of the measurements.
The card holds the listing. It carries information that looks like a listing and appeared even with web search off. An address in the named city appeared on 62% of card rows, and the trade as primary category appeared on 56%. We interpret those traits as entry conditions for appearing on the card.
Your Yelp profile shows up in answers whether you manage it or not. Yelp's rating and review count appeared most often across the card pool. We read that as a sign that a Yelp profile with a strong rating and review count gets seen in ChatGPT now. In our interpretation, Google's number takes over as a business's review footprint grows.
The prose route goes through directories and your own website. The study measured source domains and found that business sites turned retrieval into citation better than directory pages. We see presence on those directory lists as a second road into the prose answer, though the study didn't test whether a listing on any single directory raises your odds.
Making your site readable is the least expensive fix among these. In a separate census of 15,834 live independent local business websites, 23.8% were unreadable to AI crawlers. About 9% of all 15,834 sites refused AI crawlers outright through robots rules or blocking at the edge.
Track your naming rate with a band around it. The study found a 3.6% identical-list rate and 28 points of drift across three days. We'd report how often you appear across runs and phrasings, with a band around that rate, because we interpret any single check as noise.
This is the second study to come out of our AI visibility tracking work. The first examined where Claude's web search pulls its sources from, and the next two will break down phrasing results and then the card itself. If you want to know your naming rate when buyers query AI, talk to us.
