You run a bike shop, or a dental practice, or you do the books for the people who do. Someone in your town types "best place for a bike tune-up near me" into ChatGPT. That's not a hypothetical anymore: Pew Research Center reported in June that about half of American adults use an AI chatbot, and about four in ten use one to look things up.12 You've assumed, without ever checking, that the answer is Trek. The model learned the world from the internet, the internet knows the big names, so the big names win.

Lyra, our research agent, went and checked. Sixty-four prompts on August 31, 2026: eight service categories, eight cities across three metro tiers (Bellingham, Boise, Chattanooga, Dayton, Houston, Missoula, Portland, San Francisco), both engines, one day. Every prompt told the engine to search the web before answering. ChatGPT named 428 businesses across those answers and Gemini named 479.3 Of those names, 84% and 82% were local independents. National chains were 13% and 11%. Regional operators, the multi-location outfits that aren't quite either, were 4% and 5%. Gemini spent 2% of its names on platforms instead of businesses; ChatGPT spent none.

That's your number: roughly eight or nine in ten.

ChatGPTChatGPT: 84% local84%ChatGPT: 4% regionalChatGPT: 13% national13%GeminiGemini: 82% local82%Gemini: 5% regionalGemini: 11% national11%Gemini: 2% platformlocal independentsregional operatorsnational chainsplatforms
Who the assistants named. Share of all business names by class, 428 names from ChatGPT and 479 from Gemini. Table and data.

Who gets the click

A list of seven names isn't seven equal chances. The first name is the one that gets read out loud, or tapped. So the question underneath the question is who's first.

On both engines, the first business named was a local independent 88% of the time.4 Identical on ChatGPT and Gemini. Lyra flagged it as her favorite number in the whole set, and she's right, because it answers what you're actually asking. When the assistant leads with a name, it leads with a local one nearly nine times in ten.

Two numbers that sound alike and aren't

This is where it's easy to mislead yourself in either direction.

Chains are 13% of the businesses ChatGPT named and 11% of Gemini's. But 48% of ChatGPT's answers and 44% of Gemini's contain at least one chain somewhere in the list. Both figures are true at once. A chain rarely takes over an answer. It shows up as a line or two among several, and it does that in nearly half the answers.

Read the first number alone and you'll decide you're safe. Read the second alone and you'll decide chains are everywhere. The honest version: someone asking an assistant for a local service will usually see a list that's mostly local, and about half the time there'll be a national name sitting in it, sometimes more than one.

Then there's the cleanest case, answers with no chain and no regional operator at all. Those were 45% of ChatGPT's answers and 31% of Gemini's. The space between those figures and the chain-free totals (52% and 56%, by subtraction) is answers that include a regional operator, or on Gemini, a platform. Gemini pointed people to Rover three times, Find Me Gluten Free twice, and Zocdoc, Thumbtack, and Thervo once each. ChatGPT never put a platform in place of a business.

Where the chains actually are

The overall share hides the shape. Chain share by category, ChatGPT first, then Gemini:5

0%10%20%30%40%gymChatGPT gym: 34%34%Gemini gym: 33%33%plumberChatGPT plumber: 22%22%Gemini plumber: 8%8%dog boardingChatGPT dog boarding: 21%21%Gemini dog boarding: 20%20%bike shopChatGPT bike shop: 16%16%Gemini bike shop: 20%20%tax prepChatGPT tax prep: 8%8%Gemini tax prep: 5%5%gluten-free lunchChatGPT gluten-free lunch: 6%6%Gemini gluten-free lunch: 3%3%dentistChatGPT dentist: 0%0%Gemini dentist: 2%2%coffeeChatGPT coffee: 0%0%Gemini coffee: 1%1%ChatGPTGemini
National-chain share of the businesses named, by category. Gyms are the one category where a chain is a third of the list on both engines. Table and data.
  • gym: 34% / 33%
  • plumber: 22% / 8%
  • dog boarding: 21% / 20%
  • bike shop: 16% / 20%
  • tax prep: 8% / 5%
  • lunch: 6% / 3%
  • dentist: 0% / 2%
  • coffee: 0% / 1%

So the chain problem is a gym problem, a dog-boarding problem, a bike-shop problem, and on ChatGPT specifically, a plumber problem. If you run a gym, one name in three is a chain, on either engine. If you run a coffee shop or a dental practice, Gemini found one chain in each category out of sixty-odd names, and ChatGPT found none.

The brand counts confirm this and puncture a myth at the same time. ChatGPT's most-named chains were YMCA (7), Roto-Rooter (7), REI (6), Life Time (3), Planet Fitness (3), Pet Paradise (2), Camp Bow Wow (2), and Trek Bicycle (2). Gemini's were REI (5), YMCA (4), Trek Bicycle (4), Planet Fitness (3), Dogtopia (3), Camp Bow Wow (3), Roto-Rooter (3), and Equinox (2). Those are single-digit counts inside 428 and 479 names. Planet Fitness, the brand you'd assume is "the answer" for gyms, came up three times on each engine. Roto-Rooter came up seven times on ChatGPT and three on Gemini. No brand owns a category. Not one.

I'll mark one thing as my reading rather than Lyra's finding. The categories where chains cluster are the ones where a national brand is a plausible direct answer to the prompt. Ask anyone to name a gym chain and they'll have one ready. Ask for a dentist chain and watch them stall. The data is consistent with that. It doesn't prove it.

Small cities get fewer chains

I expected the reverse. My assumption was that in a small city the model would have less to go on and fall back to names it recognizes from everywhere. The data went the other way. Chain share by metro tier was 17% and 14% in large metros, 13% and 11% in mid-sized ones, and 10% and 9% in small ones. The smaller the city, the fewer the chains.

Why is a guess, and I'm not going to dress a guess up as a finding. If you're in a small city, what you can take from it is that the study found your assistants leaning more local than your counterpart's in a large metro.

The two engines don't agree with each other

One more thing that should change how you hear anyone pitching that they can get your business named inside ChatGPT or Gemini.6 Across the 64 prompts, ChatGPT and Gemini overlapped on only about a quarter of the names between them (a mean Jaccard of 0.27, if you want the statistic).7 Same prompts, same day, two different engines, and most of the businesses one named the other didn't.

Gemini names that ChatGPT also named37%ChatGPT names that Gemini also named41%
The two engines mostly named different businesses for the same prompts. Table and data.

Whatever decides which businesses get named, it isn't producing one stable list. A pitch that implies there's a single formula to crack is describing something the data doesn't show.

What we couldn't find

Here's the part I'd most like to give you and can't.

The two explanations you'll hear are that the assistants rank businesses by how often they show up in training data, and that OpenAI and Google have explained how the selection works. Lyra ran both down. Neither came back clean.

On training-data frequency: keep in mind that every prompt in this study asked for a web search first. Local Falcon, a local-search tool company, reports that when you ask Gemini for local recommendations, the answer looks a lot like live Google local results rather than something remembered from training.8 That would explain the platform answers (Thumbtack, Thervo, and Rover look like what a live search would turn up, though that's my guess) and it might explain why Gemini's plumber chain share is 8% against ChatGPT's 22%. It's one source. Lyra couldn't get a second, independent one. So it stays a contradiction in her sources, not a finding, and I'm reporting it as exactly that.

On official statements: the only Google statement Lyra could find concerns AI Overviews in Google Search, which is a different product from the Gemini assistant.9 We couldn't find anything from OpenAI about how ChatGPT selects businesses. And the line that these models were "trained on the whole internet" is contradicted by both of the sources Lyra found that describe the training data; they describe filtered or public subsets, not the whole thing.1011

The limits of the study are the limits of the study. Eight categories, three metro tiers, one day. I'm not going to tell you it holds for a florist in a city we didn't test, on a Tuesday we didn't run. It might. We didn't check.

What to hold

The first name the assistant gives is local 88% of the time, on both engines. That's the number that replaces the worry. Then look at your own category, because a gym and a coffee shop aren't in the same situation, and the table above tells you which one you're in. And when someone offers to explain the formula, ask them which of the two companies published it. As far as we could find, neither has.

Sources

  1. Pew Research Center, "Americans and AI 2026: Chatbots, Smart Devices and Views on Impact", https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/, published 2026-06-17, retrieved 2026-09-03. Quote: "About half of U.S. adults now use AI chatbots, up from a third in 2024" ↩
  2. Pew Research Center, "Americans and AI 2026: Chatbots, Smart Devices and Views on Impact", https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/, published 2026-06-17, retrieved 2026-09-03. Quote: "About four-in-ten U.S. adults say they use chatbots for information searching." ↩
  3. Data appendix, table "Who the assistants named". Aggregate tables from the study, engines.codex (ChatGPT) and engines.agy (Gemini): businesses named, share by class, first-mention share. local-vs-national-2026-08-31-summary.json, keys engines.codex.businesses = 428 and engines.agy.businesses = 479. Every scored business: local-vs-national-2026-08-31-businesses.csv. Run 2026-08-31. ↩
  4. Data appendix, table "Who the assistants named"; local-vs-national-2026-08-31-summary.json, engines.codex.first_mention_share.local = 87.5 and engines.agy.first_mention_share.local = 87.5; rounded to 88% in the text. ↩
  5. Data appendix, table "By category"; local-vs-national-2026-08-31-summary.json, by_category...share.national; for example by_category.gym.codex.share.national = 33.9 and by_category.gym.agy.share.national = 33.3. The eight prompts, the eight cities, the run date and the classification method are in the file's method block. ↩
  6. One company making this pitch, in its own words: Instant Press, "AI statistics", https://www.instantpress.co/ai-statistics, published 2026-08-10, retrieved 2026-09-04. Quote: "Instant Press helps brands get named and cited inside ChatGPT, Gemini, Perplexity and Google AI answers" ↩
  7. Data appendix, table "Cross-engine agreement"; local-vs-national-2026-08-31-summary.json, overlap.mean_jaccard = 0.267, overlap.agy_names_also_in_codex = 37.0, overlap.codex_names_also_in_agy = 41.4. ↩
  8. Local Falcon, "Where does Gemini get local business info?", https://www.localfalcon.com/blog/where-does-gemini-get-local-business-info, published 2026-06-11, retrieved 2026-09-03. One source; no independent second source found. Quote: "the typical response you get back looks an awful lot like a more conversational version of a Google 3-Pack. This is because Gemini pulls business info, including location, reviews, hours, and primary business category, directly from GBP and Google Maps." ↩
  9. MaxAEO, "How to update your business info in ChatGPT", https://maxaeo.ai/blog/update-business-info-chatgpt/, published 2026-06-11, retrieved 2026-09-03. A secondary source quoting Google; it concerns AI Overviews in Search, not the Gemini assistant. Quote: "Google states that AI Overviews use its core search index and ranking systems" ↩
  10. LLM Pulse, "What data is ChatGPT trained on?", https://llmpulse.ai/blog/chatgpt-data/, published 2026-07-09, retrieved 2026-09-03. Quote: "The single biggest ingredient is filtered text scraped from the open web. For GPT-3, OpenAI used a subset of Common Crawl" ↩
  11. Google, "Gemini overview", https://gemini.google/overview/, retrieved 2026-09-03. Quote: "since LLMs like Gemini train on the content publicly available on the internet, they can reflect positive or negative views" ↩