Who AI recommends in Los Angeles, 2026

What the assistants named

Every answer named at least one business, all 75 of them. ChatGPT named 4.2 businesses per answer on average, Gemini 4.9 and Google AI Mode 5.2.

Across their answers, ChatGPT named 102 different businesses, Gemini 121 and Google AI Mode 129. A business counts once in each place, however many of the prompts there named it.

How much they agree

To compare two assistants, the study takes the prompts both answered and, for each prompt, the businesses both named, matched by name. Over the 25 prompts each pair answered, ChatGPT and Gemini named the same business for the same prompt 24 times, ChatGPT and Google AI Mode 55 times, and Gemini and Google AI Mode 33 times.

The Jaccard index puts each pair on one scale: those shared names divided by all the businesses either one named, prompt by prompt. It was 12 percent for ChatGPT and Gemini, 31 percent for ChatGPT and Google AI Mode and 15 percent for Gemini and Google AI Mode.

ChatGPT's lists and Google AI Mode's overlapped the most, and Gemini's overlapped less with either of them than they did with each other. For every pair, fewer businesses were named by both for the same prompt than by just one assistant of the pair.

What they cite

An answer can cite sources of its own, such as Yelp, a directory or a review site. Each app shows its sources its own way, and the study recorded them as the method below says, so the counts here are what was recorded from each assistant, not a like-for-like comparison of them.

ChatGPT cited Yelp in 56 percent of its answers and a review site in 28 percent. Gemini cited a directory in 28 percent of its answers and Yelp in none of them. Google AI Mode cited Yelp in 72 percent of its answers and a map service in 56 percent.

The domains below are sources the answers cited, not the businesses they named: each assistant's listed domains cited in more than one of its answers.

• ChatGPT: yelp.com in 14 of its 25 answers, healthgrades.com in 7, webmd.com in 2.
• Gemini: zocdoc.com in 3 of its 25 answers, threebestrated.com in 2.
• Google AI Mode: yelp.com in 18 of its 25 answers, google.com in 14, zocdoc.com in 7.

Links to the businesses' own sites

A business counts as linked to its own site when an address the answer cited is that site, recognised by comparing the site's name with the business's name; nothing was looked up. How much of that the study could see turned on how each app showed its sources and how the night recorded them, so each share below stands beside its own rule, and the shares are not compared.

Of the businesses ChatGPT named, 8 percent were linked to their own sites; its source chips were kept by the domain each showed, and a chip that linked a page by that page's address. Of those Gemini named, 29 percent were; each of its source labels that gave a name was opened and read as a full address. Of those Google AI Mode named, 1 percent were, and that is a floor: its source labels that gave a business's name were not opened on this night, so such links went uncounted.

The answers also cited Yelp and map services as sources. This night recorded each of those as the site alone, with no business's own Yelp page or map place attached, so the study cannot say whether one pointed at a business the answer named, and it reports no share of businesses linked to either.

What it means for a Los Angeles business owner

The assistants did not agree on one list, so a business named by one assistant cannot count on being named by the others for the same prompt.

Yelp and directories were among the sources the answers cited, and each assistant cited a directory in some answers. The study does not show why an assistant names one business and not another, so it makes no promise about what would change an answer.

The AI visibility checker from Your Wish Is Our Code answers from this study for a category and a place, with the same limits as this page.

AI visibility checker

None of this measures results. The study counts what the assistants answered on the night it asked; it does not show whether being named brings a business more customers, and the answers can change from one run to the next.

How the study was made

Each assistant gets the same prompt for every pair of a category and a place in the plan, 25 categories and 4 places, so 100 prompts for each assistant. An automated browser session asked each prompt once, paced, in the assistant's own consumer app, one fresh chat per prompt: ChatGPT logged out, Gemini logged out and Google AI Mode logged out. The figures here are observations of AI-generated answers on the dates shown.

• The prompt, word for word: best <category> in <place>
• The places: Santa Monica, Pasadena, Sherman Oaks and Long Beach.
• The categories: med spa, dentist, orthodontist, dermatologist, chiropractor, physical therapist, plumber, electrician, HVAC company, roofer, landscaper, construction company, real estate agent, mortgage broker, personal injury lawyer, family lawyer, CPA, hair salon, nail salon, barber, gym, auto repair shop, wedding photographer, interior designer and moving company.

HVAC company and construction company are printed here as synonyms the study's AI visibility checker also accepts, in place of the trade names the plan's prompts use for those categories.

This is a partial study. A study night carries its evening's date, and its prompts can run past midnight: the first night, October 27, 2026, asked all of its prompts in the early hours of October 28, 2026, Los Angeles time. That night asked the plan's first prompts in its order: med spas, dentists, orthodontists, dermatologists, chiropractors and physical therapists in every place, and plumbers in Santa Monica. That is 25 of the 100 prompts for each of ChatGPT, Gemini and Google AI Mode, 75 answers of the 400 the plan holds, so the figures here lean on health practices and the other categories have none yet.

The study is full once every assistant in it has answered nearly all of its prompts, and more nights are planned to finish it.

Claude is in the plan, but its prompts were not asked on that night, so it has no figures here yet. Perplexity is left out on the study's own reading of its consumer terms, so the study neither asked it nor shows a figure for it.

The apps show their sources differently, and the night recorded each one's as it could. ChatGPT's source chips were kept by the domain each showed, and a chip that linked a page by that page's address. Gemini's source labels that gave a name were each opened and read as a full address. Google AI Mode's labels that gave a business's name were not opened, and on ChatGPT and Google AI Mode some labels that showed no domain were left out. The links cited per answer, the kinds of source and the shares of businesses linked to their own sites all rest on that recording, so none of them is a like-for-like comparison of the assistants, and Google AI Mode's own-site share is a floor.

Nothing was asked through an API. An assistant's app can answer differently from its maker's API, and answers change between runs, days, accounts and places, so every figure here describes one run of each prompt on the night shown.

Businesses are matched by name within each prompt, with case, punctuation and legal suffixes set aside, so a business two assistants spell differently can count as two. A business counts as linked to its own site only where the study recognised that site in an address the answer cited, and nothing was looked up.

The summary published with this page also holds, for each assistant, a share of businesses linked to a map place or a Yelp page of their own. This night recorded every such link as the site alone, so those shares read zero by how they were recorded, not by what the answers showed, and the page and its aggregates leave them out.

• Prompts answered: The plan's prompts an assistant answered, of the prompts the plan holds for each assistant.
• Answers naming a business: The answers that named at least one business, of the assistant's answers.
• Businesses named: Every business an answer named, counted once in each answer, and their mean per answer.
• Distinct businesses named: The different businesses an assistant named across its answers, each counted once in each place, matched by name with case, punctuation and legal suffixes set aside.
• Named with its own site: Of the businesses named, those whose own site was an address the answer cited, where the study recognised it. How much of that it could see turned on how each app's sources were recorded, so the shares are not compared.
• Links cited per answer: The mean of the links recorded from an answer, its sources and the businesses it linked together, as each app showed them and the study recorded them, so the means are not compared.
• Answers citing each kind of source: The answers that cited at least one link of a kind, as recorded: a map service, Yelp, a directory, a review site, a news site, a social network, a forum, a search engine or an assistant's own pages, or any other site. A business's own site, where the study recognised it, is counted under the own-site share, never as a source.
• Answers citing a listed domain: The answers that cited a domain on the study's lists of map services, Yelp, directories, review, news and social sites and forums. No business's own domain is on them.
• Businesses two assistants both named: For a pair of assistants, over the prompts both answered: the times both named the same business for the same prompt, and the share those make of all the businesses either one named, prompt by prompt (the Jaccard index).

The assistants are named in plain text as the products they are, and no endorsement by any assistant's maker is implied. The answers can be wrong, and they change. No business is named on this page or in its data, and the page ranks none.

No AI service was paid for any of it: the study used no API and no paid plan.

The data

The aggregates behind every figure on this page are published with it, one row for each assistant or pair of assistants and measure, beside the summary they come from.