---
title: "We asked ChatGPT for a plumber in 20 cities, 25 times each. It already has a favourite in yours."
description: "3,000 identical local questions to ChatGPT and Gemini in 20 cities. ChatGPT names the same business in 90%+ of answers for 49 of 60 pairs. Gemini: 5 of 60."
canonical: https://see-geo.com/blog/chatgpt-local-favourite
language: en
published: 2026-09-23
author: "SeeGeo Team"
publisher: SeeGeo
alternates:
  fr: https://see-geo.com/fr/blog/chatgpt-local-favourite
  de: https://see-geo.com/de/blog/chatgpt-local-favourite
---

# We asked ChatGPT for a plumber in 20 cities, 25 times each. It already has a favourite in yours.

> 3,000 identical local questions to ChatGPT and Gemini in 20 cities. ChatGPT names the same business in 90%+ of answers for 49 of 60 pairs. Gemini: 5 of 60.

**Quick answer: we asked ChatGPT and Gemini "Who's the best plumber in [city]?", the same for a dentist and a bakery, in 20 cities, 25 times each — 3,000 answers in one sitting. ChatGPT has a favourite: in 49 of the 60 city-and-trade pairs, one business appears in at least 90% of the 25 answers, and in 58 of 60 it appears in at least 75%. A typical ChatGPT answer names 5 businesses, and across 25 askings the whole pool is about 11. Gemini, asked the same questions, names about 43 different businesses per city and its most-named one appears in 60% of answers, median. The two engines are not reading the same web: 82% of ChatGPT's answers are business cards — rating, review count, opening status, street address — linking to the business's own website in 85% of answers, while Gemini reads Reddit, ThreeBestRated, Expertise and HomeAdvisor. For a local business, the software-brand story does not apply: there is no locked top three that you cannot join. There is a places index that has already decided, and a set of third-party lists that decide the rest.** Every number below comes from the runs; the 3,000 answers are published.

## Why this matters more for a local business than for a brand

Half of consumers now ask an AI for local recommendations. BrightLocal's Local Consumer Review Survey 2026, a panel of 1,002 US adults, found [45% had used AI to find a local business in the past year, up from 6% a year earlier](https://www.brightlocal.com/research/lcrs-ai-trust/), with ChatGPT the most used at 31%. On the other side of the same transaction, SOCi's 2026 Local Visibility Index found that across roughly 350,000 locations of 2,751 multi-location brands, [ChatGPT recommends 1.2% of locations, against a 35.9% appearance rate in Google's local three-pack](https://www.soci.ai/news/in-ai-driven-discovery-few-brands-are-chosen-most-disappear/). Demand went up sevenfold; the door stayed narrow.

[Last week we asked ChatGPT the same software-buying question 100 times](https://see-geo.com/blog/same-question-100-times) and found the top three brands locked in every answer, with everyone below them a coin flip. A software brand can read that and know where it stands. A plumber in Denver cannot, because a plumber is never a global brand. So we ran the same experiment where small businesses live.

## What we did

Twenty cities — sixteen in the US, plus Toronto, Manchester, Melbourne and Dublin — and three trades chosen because almost every town has all three and none is dominated by a national chain in an AI answer: plumbers, dentists and bakeries.

- "Who's the best plumber in Austin, Texas?"
- "Who's the best dentist in Austin, Texas?"
- "What's the best bakery in Austin, Texas?"

Sixty city-and-trade pairs, each asked 25 times to each engine on 22–23 September 2026 (UTC), from one location, through the APIs, with no conversation history and no account personalization. ChatGPT: `gpt-4.1` with the web-search tool available; it searched in 1,500 of 1,500 answers. Gemini: `gemini-flash-latest` with Google Search grounding offered; it used it in 1,356 of 1,500 — the opposite of its behaviour on the software questions last week, where it used search once in 300.

A business counts as **named** in an answer when it appears as a heading, a list item, a table row or a bold lead — the places an answer puts its picks. Names were normalised (punctuation, "LLC", "Inc." removed) and a shorter form was folded into a longer one within the same city when the longer one was at least as frequent, so "Radiant Plumbing" and "Radiant Plumbing & Air Conditioning" count as one business. Every heading that looked like a business was reviewed by hand; generic headings ("Tips for hiring a plumber", "Availability") were excluded. Full rules at the end.

## ChatGPT already has a favourite in your city

In 49 of 60 city-and-trade pairs, one business appeared in at least 90% of the 25 answers. In 58 of 60 it appeared in at least 75%. In no pair was the most-named business below 50%. The median answer names 5 businesses; across all 25 askings of a pair, the pool of businesses ever named has a median size of 11.

| City | ChatGPT names most often | share | distinct businesses | Gemini names most often | share | distinct businesses |
|---|---|---|---|---|---|---|
| Austin | Abacus Plumbing Air Conditioning & Electrical | 100% | 7 | Clarke Kent Plumbing | 84% | 49 |
| Boise | Proficient Plumbing Service | 100% | 12 | Goodson Plumbing Services | 40% | 43 |
| Charlotte | Queen City Plumbing | 100% | 13 | Wyattworks Plumbing | 60% | 48 |
| Columbus | Parson Plumbing And Drains | 60% | 18 | Drainmaster | 72% | 30 |
| Denver | Brothers Plumbing Heating & Electric | 96% | 8 | Hyper Flow Service Company | 72% | 52 |
| Dublin | Plumbers Dublin | 100% | 5 | Dublin Area Plumbers | 92% | 38 |
| Kansas City | A 1 Sewer & Septic | 96% | 10 | Poor Johns Plumbing | 72% | 39 |
| Madison | Bartow Plumbing | 100% | 11 | Triggs Plumbing | 32% | 33 |
| Manchester | Manchester Plumbers | 92% | 14 | Aquarius Plumbing | 92% | 63 |
| Melbourne | Butler Plumbing Pty | 100% | 8 | Fix It Right Plumbing | 68% | 91 |
| Minneapolis | Gray Duck Plumbing | 88% | 9 | Ron The Sewer Rat | 48% | 44 |
| Nashville | Morton Plumbing Heating & Cooling | 100% | 7 | Hiller Plumbing Heating Cooling & Electrical | 64% | 39 |
| Omaha | Plumbrite | 100% | 9 | Big Birge Plumbing | 44% | 35 |
| Pittsburgh | Restano Heating Cooling & Plumbing | 88% | 14 | Chris Driscoll Plumbing | 68% | 69 |
| Portland | Pipeline Plumbing & Drains | 100% | 6 | Meticulous Plumbing | 76% | 50 |
| Richmond | Harrys Plumbing | 92% | 19 | Robinsons Plumbing Service | 84% | 50 |
| Sacramento | Millennium Plumbing Specialist | 100% | 18 | Sacramento Plumbing Solutions | 60% | 54 |
| Tampa | True Plumbers & Ac | 96% | 8 | Everydayplumber Com | 40% | 33 |
| Toronto | Drainworks Plumbing | 96% | 10 | Leaside Plumbing And Heating | 48% | 52 |
| Tucson | Imperial Plumbing | 100% | 9 | Cals Plumbing | 64% | 43 |

The same shape holds for dentists (favourite in at least 90% of answers in 14 of 20 cities, median pool 11.5) and bakeries (18 of 20, median pool 14). The weakest lock in the whole set was Parson Plumbing And Drains in Columbus, still named in 60% of answers.

The first name in the list is less fixed than the presence: the modal first-named business led 72% of answers, median across pairs. Order moves; membership does not. Two consecutive ChatGPT answers to the same question shared 58% of their businesses on average.

## Where ChatGPT's list comes from

This is the finding that changes what a local business should do. 82% of ChatGPT's 1,500 local answers are not prose recommendations at all. They are cards:

> **Abacus Plumbing, Air Conditioning & Electrical**
> Open now · Plumber · 4.7 (508 reviews)
> 708 S Lamar Blvd G, Austin, TX 78704

Rating, review count, opening status, street address, then a sentence. That is the shape of a places record, not of a web page the model read and summarised. In 85% of answers the card links to the business's own website; where a business has no site on record, the link is a Google Maps search for its name instead (15% of answers, 1,749 such links). Of the 7,581 links in ChatGPT's local answers, 527 carry tracking parameters that the business itself attached to the website link in its Google Business Profile — the URL travelled from the profile into the answer intact.

Put plainly: for a local question, ChatGPT is not surfing listicles the way it does for software. It is querying a places index that looks and behaves like a business-listing database, choosing a handful of entries with high ratings and large review counts, and formatting them as cards. The favourite in each city is the entry that index returns first, every time. Which is why the lock is so tight, and why it is a very different lock from the software one: it is not "the brands every listicle names"; it is "the listing the index ranks first for this query".

## Gemini reads a different web

Gemini, given Google Search grounding on every call, used it in 1,356 of 1,500 answers and produced something else entirely. Its answers are essays with sections — "best independent and local favourites", "best for emergencies" — naming 6 businesses per answer, median, from a pool of 43 per city across 25 askings. Its most-named business appeared in 60% of answers, median, and in only 5 of 60 pairs did any business reach 90%. Two consecutive Gemini answers shared 16% of their businesses.

And its sources are the lists. Across 1,356 grounded answers, Gemini cited 887 distinct domains, a directory or review site in 91% of answers, and one domain above all others: Reddit, cited in 78% of its grounded answers. After Reddit come facebook.com (315), forbes.com (129), healthgrades.com (106), bbb.org (92), angi.com (82), homeadvisor.com (75). In 124 of 1,500 answers Gemini named no business at all — 87 of them bakery questions, where it often described what to look for instead of where to go.

| City | ChatGPT names most often | share | distinct businesses | Gemini names most often | share | distinct businesses |
|---|---|---|---|---|---|---|
| Austin | Breeze Dental | 68% | 17 | Toothbar | 84% | 41 |
| Boise | Summit Dental Group | 100% | 9 | Summit Dental Group | 76% | 48 |
| Charlotte | Charlotte Dentistry | 84% | 14 | Caldwell Bills Petrilli & West | 96% | 91 |
| Columbus | Grandview Dental Care | 88% | 9 | Firouzian Dentistry | 76% | 63 |
| Denver | Metropolitan Dental Care | 100% | 15 | Metropolitan Dental Care | 88% | 69 |
| Dublin | Truly Dental | 92% | 13 | Seapoint Clinic | 84% | 64 |
| Kansas City | Nkc Dental | 100% | 11 | Evers & Gardner Dental | 88% | 106 |
| Madison | Affiliated Dentists S C | 96% | 17 | Total Care Dental | 28% | 42 |
| Manchester | Kissdental Manchester | 100% | 9 | Ringway Dental | 76% | 54 |
| Melbourne | Smile Solutions | 96% | 9 | Smile Solutions | 60% | 61 |
| Minneapolis | J & D Dental | 100% | 8 | Park Dental | 88% | 78 |
| Nashville | Thomas Nabors | 100% | 12 | Belle Meade Family Dentistry | 80% | 74 |
| Omaha | Premier Dental | 100% | 8 | Premier Dental | 60% | 54 |
| Pittsburgh | Edgewood Dental | 100% | 11 | Atlas Dental Specialists | 96% | 68 |
| Portland | Cedar Creek Dental | 100% | 7 | Belmont Family Dentistry | 88% | 90 |
| Richmond | River Run Dental | 80% | 29 | Virginia Family Dentistry | 60% | 52 |
| Sacramento | Midtown Dental | 100% | 10 | Midtown Dental | 56% | 63 |
| Tampa | Tampa Palms Dentistry | 96% | 12 | Mcilwain Dental Specialists | 52% | 83 |
| Toronto | Toronto Beach Dental | 88% | 18 | Dr Judy Sturm & Associates | 96% | 93 |
| Tucson | Casas Adobes Dentistry | 88% | 22 | Martin Taylor Dentistry | 60% | 91 |

So the two engines with the largest share of local AI queries disagree not on which business is best, but on what kind of evidence counts. ChatGPT trusts a listing. Gemini trusts a list.

## What a local business should do with this

**Your Google Business Profile is now an AI data source, and ChatGPT's card is built from it.** The rating, the review count, the hours, the address and the website link in an answer are profile fields. A profile with a wrong website, no hours, or a handful of reviews is a card that the index ranks below the one with 508 reviews. This is unglamorous work — categories, hours, the website field, review volume — and it is the whole of ChatGPT local visibility today. We wrote the [local business guide to AI recommendations](https://see-geo.com/blog/local-business-ai-recommendations-guide) around exactly this; the data above is why.

**Your website still has to be readable, because it is what the card links to.** 85% of ChatGPT's cards send the reader to the business's own site. A site that loads slowly, hides its services behind JavaScript, or has no plain statement of what the business is and where wastes the click the index just gave it. That is what our [free audit](https://see-geo.com/) measures.

**For Gemini, get onto the lists — starting with Reddit.** The local subreddit's "who's a good plumber" thread, then the review directories above, decide Gemini's answer, and a business that is not on them is not in the pool of 43. Ask for inclusion where there is a process; answer the Reddit question where there is a thread.

**Measure with repeated sampling, and measure each engine separately.** One answer from ChatGPT tells you almost everything, because its list barely moves. One answer from Gemini tells you almost nothing, because its pool is more than forty deep. A visibility number that averages the two describes neither. SeeGeo's tracking asks each engine on its own, repeatedly, and reports rates per engine for this reason.

## Frequently asked questions

**Does ChatGPT recommend the same local business every time?**
Nearly. In 49 of 60 city-and-trade pairs, one business appeared in at least 90% of 25 repeated answers, and the median answer names 5 businesses from a pool of about 11. The order within the list moves; the membership barely does.

**Where does ChatGPT get its local business recommendations?**
From a places index, not from web pages. 82% of its answers in this study are cards with a rating, review count, opening status and address, linking to the business's own website in 85% of answers or to a Google Maps search when no site is on record. 527 of the links still carry the tracking tags businesses add to the website field of their Google Business Profile.

**Why does Gemini give a different list than ChatGPT?**
Because it reads different evidence. Gemini used Google Search grounding in 1,356 of 1,500 answers and cited directories and ranking sites — ThreeBestRated, Expertise, HomeAdvisor, Reddit — naming a median of 43 businesses per city across 25 askings against ChatGPT's 11.

**How do I get my business recommended by ChatGPT?**
Make the listing the index ranks first: a complete Google Business Profile with the right category, hours, a working website link and a large, recent, genuine review count. Then make the website that link opens readable to a machine. There is no listicle to get onto for ChatGPT's local answers; there is a record to complete.

**Can I see the raw answers?**
Yes. All 3,000 answers, the extraction rules and the per-city statistics are published at [github.com/Tunisian-Aaron/ai-visibility-study-2026](https://github.com/Tunisian-Aaron/ai-visibility-study-2026) under CC BY 4.0. Businesses are named because the engines named them; no business was audited or contacted for this study.

## Full methodology

*Published verbatim; runs conducted 22–23 September 2026 (UTC), from a single location, through the engines' APIs.*

**Questions.** Sixty: three trades × twenty cities. Trades: plumber ("Who's the best plumber in {city}?"), dentist ("Who's the best dentist in {city}?"), bakery ("What's the best bakery in {city}?"). Cities: Austin TX, Denver CO, Portland OR, Nashville TN, Charlotte NC, Columbus OH, Boise ID, Tucson AZ, Omaha NE, Tampa FL, Sacramento CA, Kansas City MO, Minneapolis MN, Pittsburgh PA, Richmond VA, Madison WI, Toronto, Manchester UK, Melbourne, Dublin. One phrasing per trade; paraphrase variance is not measured.

**Engines and settings.** ChatGPT: OpenAI Responses API, model `gpt-4.1`, `web_search_preview` tool available on every call, provider default temperature, no system prompt, no history; whether the model searched is read from the response's `web_search_call` items. Gemini: `gemini-flash-latest` via the Generative Language API, `google_search` grounding offered on every call; an answer counts as grounded when grounding metadata is returned. Four concurrent requests per engine; each call retried up to three times on transport error; no call failed after retry. 25 runs per pair per engine, 3,000 answers.

**Extraction.** A business is *named* when it appears in a heading, a numbered or bulleted item, a table row, or a bold-led line. Names are normalised (case, punctuation, "LLC/Inc/Ltd/Co/DDS/PC" removed) and, within a city-and-trade pair, a shorter name is folded into a longer name that starts with it when the longer one is at least as frequent. Headings matching a stop-list of generic phrases (tips, availability, pricing, how to choose, …) are excluded; the residue of unmatched headings was reviewed by hand and the stop-list extended before the final pass. *Named first*: the first named business in reading order. *Overlap*: mean Jaccard similarity of named-business sets over all pairs of answers in a cell. *Card*: an answer containing a "rating (N reviews)" line. *Own-site link*: a cited URL whose domain is neither a directory nor Google/Facebook. Intervals, where given, are Wilson 95%.

**Limitations.** One day, one location, one phrasing per trade, English only, two engines with one model each. Results describe these runs. Extraction is heuristic: a business named only in prose is not counted, and a fold rule can, in principle, merge two businesses whose names share a prefix; the per-city tables were checked by eye for that. "Places index" is our description of the observed output shape and link provenance, not a statement about OpenAI's internal architecture. We did not verify ratings or review counts against any platform, and we did not measure answer quality.