Who does AI pick?

We asked eight AI models questions like "What is the best coffee shop in Chicago?" for six kinds of local business in five US cities, twice on November 1, 2026. We recorded which businesses each answer named.

Key findings

Numbers are from the first asking (238 answers, of which 208 came from the seven models with web search) unless the finding says otherwise.

1. For each question, the eight models together named a median of 32 different businesses (range 18 to 50 across the 30 questions).
2. Most businesses came from a single model: of the 983 businesses named across the 30 questions (each counted once per question), 564 (57 percent) were named by only one of the eight models.
3. Most questions still had a business named by a majority of the models: in 28 of the 30 questions at least 5 of the 8 models named the same business, and in 21 of the 30 the most-named business was named by at least one model from each of the study's three separate web search systems (Azure's, Anthropic's and Perplexity's).
4. Two models answering the same question mostly listed different businesses: on average, only 17 percent of the businesses either model named were named by both (across all 28 model pairs and 30 questions).
5. The models often disagreed about which business to name first: the eight answers to a question had a median of 5 different first-named businesses (30 questions).
6. Asked the same question again a few minutes later, the models named 1,135 of the 1,880 businesses they had named the first time (60 percent), from 80 percent for Anthropic's Claude (256 of 319) to 27 percent for Google's Gemini 3.5 Flash-Lite, which answered without web search (69 of 252).
7. National and multi-state chains made up 192 of the 1,880 times a model named a business (10 percent), but 115 of 364 in gym answers (32 percent) and 4 of 283 in hair salon answers (1 percent).
8. The websites cited most often were ranking and directory sites: bestinhood.com and trustanalytica.org each appeared in 38 of the 208 answers that cited sources, Yelp in 32, Reddit in 2 and Google Maps in none.
9. Local publications and blogs appeared in 104 of the 208 answers that cited sources (50 percent), and the website of a business the answer named appeared in 76 of the 208 (37 percent).
10. Google's Gemini 3.5 Flash-Lite, which answered from training without web search, overlapped least with the other models: 181 of the 252 businesses it named (72 percent) were named by none of the seven web-search models for the same question.

What this means for a local business

The sites the models cited most

The ten websites cited in the most answers in the first asking. Each count is how many of the 208 answers that cited sources included the site. The Method section below explains the types.

Who the models named, city by city

The three businesses named by the most models for each question, with the count from the first asking and then from the second. "Chain" marks a national or multi-state chain. When several businesses tie for the third row, we list the one named by more models in the second asking and say how many others tied.

Being named by more models says nothing about a business's quality or service.

Chicago

Business	First asking	Second asking
Coffee shop
Metric Coffee	named by 7 of 8 models	8 of 8
Sawada Coffee	named by 6 of 8 models	6 of 8
Intelligentsia chain	named by 6 of 8 models	5 of 8
Hair salon
Sparrow	named by 5 of 8 models	6 of 8
GC By George Charles	named by 4 of 8 models	4 of 8
Sine Qua Non	named by 4 of 8 models	4 of 8
Barbershop
The Belmont Barbershop	named by 6 of 8 models	5 of 8
Joe's Barbershop	named by 5 of 8 models	5 of 8
Elegance Hair Salon & Barbershop	named by 4 of 8 models	4 of 8
2 other businesses were also named by 4 of 8 models.
Gym
Midtown Athletic Club chain	named by 7 of 8 models	6 of 8
Equinox chain	named by 6 of 8 models	5 of 8
East Bank Club	named by 5 of 8 models	7 of 8
3 other businesses were also named by 5 of 8 models.
Clothing boutique
Una Mae's	named by 5 of 8 models	7 of 8
Alice & Wonder	named by 4 of 8 models	4 of 8
Bronzeville Boutique by Lady Mocha	named by 4 of 8 models	4 of 8
3 other businesses were also named by 4 of 8 models.
Brunch spot
Lula Cafe	named by 8 of 8 models	8 of 8
The Publican	named by 5 of 8 models	5 of 8
Mon Ami Gabi chain	named by 4 of 8 models	4 of 8
2 other businesses were also named by 4 of 8 models.

Denver

Business	First asking	Second asking
Coffee shop
Corvus Coffee Roasters	named by 7 of 8 models	5 of 8
Huckleberry Roasters	named by 6 of 8 models	7 of 8
Little Owl Coffee	named by 4 of 8 models	5 of 8
1 other business was also named by 4 of 8 models.
Hair salon
303 Salon	named by 7 of 8 models	6 of 8
Three Cutters	named by 4 of 8 models	4 of 8
Grand Salon & MedSpa	named by 4 of 8 models	2 of 8
Barbershop
Blank Barbers	named by 6 of 8 models	5 of 8
Semion Barbershop	named by 5 of 8 models	6 of 8
Alfanis Barber Shop	named by 5 of 8 models	2 of 8
Gym
The Athletic Club	named by 6 of 8 models	5 of 8
Denver Athletic Club	named by 4 of 8 models	3 of 8
Club Form Denver	named by 3 of 8 models	3 of 8
1 other business was also named by 3 of 8 models.
Clothing boutique
Inspyre Boutique	named by 5 of 8 models	6 of 8
Rockmount Ranch Wear	named by 4 of 8 models	4 of 8
Patterns & Pops	named by 4 of 8 models	3 of 8
Brunch spot
Safta	named by 5 of the 6 models that answered	5 of the 7 that answered
Snooze, an A.M. Eatery chain	named by 5 of the 6 models that answered	5 of the 7 that answered
Onefold	named by 4 of the 6 models that answered	6 of the 7 that answered
1 other business was also named by 4 of the 6 models that answered.

Nashville

Business	First asking	Second asking
Coffee shop
Crema Coffee Roasters	named by 8 of 8 models	8 of 8
Barista Parlor	named by 5 of 8 models	4 of 8
Frothy Monkey chain	named by 3 of 8 models	4 of 8
5 other businesses were also named by 3 of 8 models.
Hair salon
Dandelion Salon	named by 6 of 8 models	6 of 8
Style House Salon	named by 5 of 8 models	4 of 8
Element Salon	named by 4 of 8 models	6 of 8
Barbershop
Scout's Barbershop	named by 6 of 8 models	6 of 8
Hawkins Barbershop	named by 3 of 8 models	3 of 8
Collins & Co.	named by 3 of 8 models	2 of 8
2 other businesses were also named by 3 of 8 models.
Gym
YMCA chain	named by 8 of 8 models	6 of 8
Carbon Performance	named by 6 of 8 models	6 of 8
QNTM Fit Life	named by 5 of 8 models	6 of 8
2 other businesses were also named by 5 of 8 models.
Clothing boutique
Vinnie Louise	named by 7 of 8 models	5 of 8
Emerson Grace	named by 5 of 8 models	5 of 8
H. Audrey	named by 5 of 8 models	5 of 8
1 other business was also named by 5 of 8 models.
Brunch spot
Biscuit Love chain	named by 6 of 8 models	5 of 8
The Butter Milk Ranch	named by 5 of 8 models	6 of 8
Noko	named by 4 of 8 models	4 of 8

Seattle

Business	First asking	Second asking
Coffee shop
Espresso Vivace	named by 6 of 8 models	7 of 8
Milstead & Co.	named by 6 of 8 models	3 of 8
Analog Coffee	named by 5 of 8 models	5 of 8
2 other businesses were also named by 5 of 8 models.
Hair salon
Antonio Salon	named by 5 of 8 models	4 of 8
Vann Studio Salon	named by 4 of 8 models	5 of 8
Fix Salon	named by 4 of 8 models	4 of 8
1 other business was also named by 4 of 8 models.
Barbershop
Assembly Barbershop	named by 6 of 8 models	7 of 8
Squire Barbershop	named by 6 of 8 models	6 of 8
The Scotch Pine	named by 6 of 8 models	5 of 8
Gym
Rain City Fit	named by 7 of 8 models	7 of 8
Seattle Athletic Club	named by 6 of 8 models	7 of 8
Ballard Health Club	named by 5 of 8 models	5 of 8
1 other business was also named by 5 of 8 models.
Clothing boutique
Glasswing Shop	named by 7 of 8 models	7 of 8
Bon Voyage Vintage	named by 5 of 8 models	5 of 8
Pipe and Row	named by 5 of 8 models	4 of 8
Brunch spot
Glo's	named by 6 of 8 models	5 of 8
Tilikum Place Cafe	named by 5 of 8 models	6 of 8
Portage Bay Cafe	named by 5 of 8 models	3 of 8

Phoenix

Business	First asking	Second asking
Coffee shop
Songbird Coffee & Tea House	named by 4 of 8 models	8 of 8
Moxie Coffee Co.	named by 4 of 8 models	4 of 8
Cartel Coffee Lab	named by 3 of 8 models	5 of 8
1 other business was also named by 3 of 8 models.
Hair salon
Salon D'Shayn	named by 6 of 8 models	7 of 8
The Root Salon	named by 5 of 8 models	5 of 8
Salon Surreal	named by 5 of 8 models	4 of 8
Barbershop
Uptown Barbershop	named by 6 of 8 models	4 of 8
Ace of Fades Executive Barbershop	named by 5 of 8 models	5 of 8
The Local Barber & Shop	named by 5 of 8 models	5 of 8
Gym
Mountainside Fitness	named by 7 of 8 models	7 of 8
Life Time chain	named by 7 of 8 models	6 of 8
EōS Fitness chain	named by 5 of 8 models	6 of 8
Clothing boutique
Frances	named by 7 of 8 models	6 of 8
Bunky Boutique	named by 6 of 8 models	5 of 8
Phoenix General	named by 4 of 8 models	5 of 8
1 other business was also named by 4 of 8 models.
Brunch spot
Lon's at the Hermosa Inn	named by 4 of 8 models	5 of 8
Valentine	named by 4 of 8 models	5 of 8
Brunch & Sip	named by 4 of 8 models	4 of 8

Method

When. Sunday, November 1, 2026, between 6:08 pm and 6:23 pm US Pacific time (November 2, 02:08 to 02:23 UTC). That was evening on November 1 in all five cities. The first asking ran from 02:08 to 02:12 UTC and the second from 02:13 to 02:16 UTC; we retried failed calls until 02:23 UTC.

Questions. One question per city and category, sent exactly as written below with no other instructions, for 30 questions in all:

• What is the best coffee shop in {city}?
• What is the best hair salon in {city}?
• What is the best barbershop in {city}?
• What is the best gym in {city}?
• What is the best clothing boutique in {city}?
• What is the best brunch spot in {city}?

Cities. Chicago, Denver, Nashville, Seattle and Phoenix.

Models. Eight models, each asked through a developer API with the same setup the AUTX AI Visibility Report uses:

• Anthropic's Claude Sonnet 5.5, through Anthropic's API, with Anthropic's web search.
• OpenAI's GPT-6.1 Sol and GPT-6 Luna, through Microsoft Azure, with Azure's web search tool.
• xAI's Grok 4.7, Meta's Muse Spark 1.3 and DeepSeek's DeepSeek V4 Pro, through Microsoft Azure, with the same Azure web search tool.
• Google's Gemini 3.5 Flash-Lite, through the Gemini API, without web search: it answered from what it learned in training and cited no sources.
• Perplexity's Sonar, through Perplexity's API, with Perplexity's own search. It received only Perplexity's standard instruction on how to format citations.

Askings. Each of the 30 questions went to each of the 8 models twice, a few minutes apart (the first and second asking): 480 planned answers. Of the 480 first attempts, 19 failed: 11 were rate-limit errors on GPT-6 Luna, and 8 were answers that Microsoft Azure's content filter cut to the single line "I'm sorry, but I cannot assist with that request." (7 from GPT-6 Luna, 1 from GPT-6.1 Sol). We retried failed calls with the same settings, up to three more times, for 506 model calls in all. We collected 477 of the 480 answers. The 3 missing answers are all for the Denver brunch question, which Azure's filter blocked on every try for GPT-6.1 Sol in the first asking and for GPT-6 Luna in both askings.

Reading the answers. Claude Sonnet 5.5, without web search, listed the businesses each answer put forward as recommendations, in the order they appeared. It did not count neighborhoods, streets, shopping centers, markets, publications, review sites, apps, individual people, products, or businesses mentioned only as a supplier, an owner, a comparison or a closed place. Every extracted name appears in the text of its answer. We checked 48 answers by hand (10 percent, six per model, picked at random): 393 names, none wrongly added and no clear misses. We matched names within each city and category at the business level, so a name written with and without words such as "Coffee", "Co." or "Salon", or with a branch or neighborhood added, counts as one business. We merged 55 remaining spelling variants by hand.

Chains. A chain is a national or multi-state business, franchise or network, meaning it has locations in two or more US states; the YMCA counts as a chain. Everything else counts as local, including local businesses with several locations in one state and businesses we could not identify. Claude Sonnet 5.5 made a first pass. We reviewed by hand every business it marked as a chain and every business named four or more times, and we checked the uncertain high-count cases against the businesses' published locations.

Sources. We reduced each cited web address to its main domain (for example, yelp.com) and gave it a type: review site, ranking or directory site, local publication or blog, national guide, business website, tourism site, booking site and so on. We reviewed the types by hand. "Local publications and blogs" are outlets focused on one city or region, such as city magazines, alt-weeklies, local TV sites and local blogs. City editions of national guides, such as Eater and Time Out, count as national guides.

Spend. $20.52 in total: $15.81 for the 506 model calls, $0.08 for 3 diagnostic calls, and $4.64 for 529 calls that extracted names and classified businesses and websites.

Limits

• A snapshot. All answers come from one evening, November 1, 2026, with two askings a few minutes apart. Answers change between askings: in the second asking, the models again named 60 percent of the businesses they had named in the first.
• One wording. We used one wording per category ("What is the best ... in {city}?"). Other wordings, such as "near me", "cheap" or "open now", can bring different answers.
• Developer APIs, not consumer apps. We asked each model through a developer API. The study does not measure ChatGPT the app, Google Search, Google AI Mode or any other consumer product, where account settings, location and app features can change answers.
• Models change. Model versions and the web pages their search tools find change often. The results describe these model versions on November 1, 2026.
• Shared search. Five of the eight models (OpenAI's GPT-6.1 Sol and GPT-6 Luna, xAI's Grok 4.7, Meta's Muse Spark 1.3 and DeepSeek V4 Pro) used the same Azure web search tool, so part of their agreement reflects shared search results. Grok 4.7, Muse Spark 1.3 and DeepSeek V4 Pro overlapped with each other more than any other models did.
• No web search for one model. We asked Gemini 3.5 Flash-Lite without web search, as in AUTX's paid report, so its answers reflect what it learned in training rather than current web pages.
• Missing answers. Of the 480 answers, 3 are missing, all for the Denver brunch question, because Microsoft Azure's content filter blocked OpenAI's models on every try.
• Not a ranking of quality. Being named by more models says nothing about a business's quality or service. We recorded names as the models wrote them and did not check that each business exists, is open or matches the details a model gave.
• Counting choices. Branches of one business count as one business. Chain status reflects our classification, and we may have misclassified a few small businesses with locations in two states. A business in a suburb counts toward the city it was named for.

The data

For reporters and researchers: data.json (JSON, 391 KB) lists the 30 questions and the eight models and, for each answer, the businesses it named in order, with a chain flag, and the web domains it cited. It leaves out the full answer texts. The data is free to use under the Creative Commons Attribution 4.0 license (CC BY 4.0): please credit AUTX and link to this page.