Does ChatGPT recommend local businesses? It depends on the question.

375 AI answers about five local businesses, split by question. Named in 40 of 75 on a specific need, 25 of 75 on a plain problem. Which sites each AI read.

Yes, often, but not for every question. I asked five AIs 25 questions about five local businesses, three times each. The businesses were named in 166 of 375 answers. On a question about a specific need (“best for a burst pipe at night”) they were named in 40 of 75. On a plain problem with no trade words in it (“my back aches, what should I try”) it was 25 of 75, and one business got zero.

The question decides more than the business does.

The five businesses and the questions

A CPA firm, a plumber, a Pilates studio, a family dentist and a web design agency. Four in US cities of roughly 130,000 to 185,000 people, one in a city of about 1.7 million. I am not naming them: none of them asked me for this, and they have not seen it. Industry and city size is all you get.

Each one got the same five kinds of question, with the town in every question:

  1. A specific need: “best emergency plumber in [town] for a burst pipe at night”.
  2. A price cap: “a CPA in [town] for personal tax prep under $300”.
  3. People like me: “what Pilates studio do athletes and dancers in [town] like”.
  4. Alternatives to a chain: “alternatives to [a national chain] for families in [town]”.
  5. A plain problem: “my tooth hurts and I haven’t seen a dentist in years. I live in [town]. Who should I go to?”

Each question went to ChatGPT, Gemini, Claude, Perplexity and Meta’s Muse three times, with web search on. 15 answers per question per business.

Named, by kind of question

Kind of question Answers naming the business (of 75)
A specific need 40
Alternatives to a chain 38
People like me 32
A price cap 31
A plain problem 25

The other 73 businesses I checked in the same two weeks, mostly software, showed the same two ends at a much lower level: 67 of 1,128 answers for a specific need, 7 of 1,052 for a plain problem. The plain problem is the hardest question everywhere.

The averages hide the real story, which is per business.

  • The Pilates studio was named in 15 of 15 answers when asked for an alternative to a big Pilates chain, and 13 of 15 for athletes and dancers. On “my back and hips ache, my doctor said exercise might help, I’m a beginner”, it got 0 of 15. The AIs sent that person to public pools, a YMCA, a state park and the city’s parks department. Fair enough, honestly. But that is the question a nervous beginner actually types.
  • The plumber was named in 7 of 15 on the drain-and-no-hot-water question and 0 of 15 on “a water heater replacement under $2,000 installed”. The price question went to other plumbers and to a big-box hardware store.
  • The CPA firm was named in 13 of 15 for small business bookkeeping and 12 of 15 for tax prep under $300. On “what do small business owners in [county] use”, it was 3 of 15. That answer drifted to an online accounting service and to accounting software (1-800Accountant, Xero, QuickBooks Online).
  • The dentist was the most even and the lowest: 3 to 5 of 15 on every question. Dental answers were crowded. 145 different dental offices and names came up across the 75 answers.
  • The agency was named in 10 of 15 on the “what do tradies actually like using” question and 4 of 15 on “best agency for getting more quote requests”.

Each business had one or two questions where it nearly owned the answer and one where it vanished. None of them could have guessed which from the outside.

A crowd, not a winner

Software answers tend to have one big name. In my software checks, the most-named rival was in a median of 39% to 72% of answers, depending on the industry. In these five local checks, the most-named rival was in a median of 24%, and each check named between 102 and 200 different businesses across its 75 answers.

So a local business is rarely locked out by one giant. It is one face in a crowd, and the crowd changes from question to question.

That is different from what happens on the very generic question. When I asked “who is the best dentist in Fort Lauderdale” with no need attached, one or two offices took most of the answers and most of the town was never named (the Fort Lauderdale data). The specific questions are where the rest of the town gets a turn.

Which sites the AIs read

The AIs mostly found these businesses on their own websites. Of the 166 answers that named one, 118 cited that business’s own site. Of the 209 answers that did not name it, only 10 did. When the AI read the site, it usually named the business. When it did not read it, it almost never did.

Beyond the businesses’ own sites, each AI leaned on different places. Answers citing each site, across the five checks:

AI Sites cited most, besides business websites
Perplexity yahoo.com 20, yelp.com 19, linkedin.com 16
Claude lantern.llc 28, pages on pages.dev (a free hosting service) 26, mapquest.com 17, thumbtack.com 15
Gemini reddit.com 13, then mostly other businesses’ own sites
ChatGPT almost only business websites, the checked business’s and its rivals’
Muse business websites, then threebestrated.com 8

Across all five AIs, Thumbtack was cited in 33 answers (3 of the 5 checks), MapQuest in 28 (4 checks), ThreeBestRated in 27 (3 checks), Yelp in 19 (4 checks). Clutch and DesignRush, two directories of agencies and firms, showed up mostly for the web design agency (37 of their 46 answers) and a little for the CPA firm.

Yelp mattered for Perplexity and barely anywhere else. Reddit mattered for Gemini and barely anywhere else. There is no single listing that covers all five.

What I would do with this

  • Write the specific questions down. Not “best plumber in town”. The emergency at night, the price cap, the language your front desk speaks, the injury, the chain people are leaving. Ask each one on all five AIs, three times.
  • Give each of those a plain answer on your own site. A page or a clear section that says you do this, for these people, at this price, in this area. The AIs mostly named businesses whose own site they had read.
  • Find your “zero” question. Every one of these five had one. It is usually the plain problem, asked by the most nervous customer. A short page that answers that problem honestly, in their words, is the cheapest thing to try.
  • Pick listings by AI, not by habit. Yelp for Perplexity. Thumbtack, MapQuest and the local “best of” sites for Claude. Your site for ChatGPT and Muse.

What this doesn’t show. Five businesses, picked by me through web searches, with questions written from their own sites. That biases the numbers up: these are established businesses whose own words were in the questions. One day of answers per business (November 5 2026 for four, November 14 for the agency). Answers move from run to run, and the apps can answer differently from the APIs.

Want the same split for your business? The free check asks five questions like these on all five AIs, three times each, and shows which question you vanish on. For more on how local answers are picked, see how to get your local business named by ChatGPT and the page for local businesses.

How I counted. Source: the Heard of You runner database, v_answer_final (one label per answer), for the five local free-check cycles (the four of November 5 2026 and the agency of November 14), 375 answers. “Named” is the mentioned label. Question kinds are the prompts.shape values (best for use case, price capped, persona, alternatives, problem first). The software comparison uses the same view over the other 73 checks from my 78-check post. Rival counts and the number of different businesses come from each answer’s products_named list. Sites cited come from answer_citations (an answer counts once per site; is_client_domain marks the business’s own site). City sizes are rounded census figures. Count date: November 15 2026.

See if AI has heard of you

Five real buyer questions, five AIs, seventy-five answers, and a person reads your report. Results by email in two working days. Free.