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.

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

• 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.

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.

The next day the same tool ran against the site of an agency that sells AI search optimization. GPTBot, ClaudeBot, and Google-Extended, all three disallowed in the file it serves.¹ We aren't naming them. We're writing this because we didn't know about ours either until something went and looked.

You wrote a robots.txt. It's in the repo, it allows the crawlers you want, and you've probably looked at it more than once this year. Here's the thing you may not have done: fetched it from outside, the way a crawler does, and read what came back.

If your site is behind Cloudflare, that fetch can return a different file. Lyra, our research agent, went looking for how, and everything she could find came from Cloudflare's own documentation. That's worth saying up front. This is Cloudflare describing Cloudflare, so read it as the vendor's account of its own feature, and check your own site rather than taking either of us on faith.

What the managed file does

Cloudflare offers a feature it calls managed robots.txt. When it's on, Cloudflare checks whether your server already serves a robots.txt, judged by an HTTP 200 response, and if it does, Cloudflare prepends its own directives ahead of yours and sends both back as a single file.² If your server has no robots.txt, Cloudflare creates one.³

The result is a file with two authors. Your directives are still in there, untouched. Cloudflare's sit above them. What Cloudflare's block says, in Cloudflare's own blog post about the feature, is a request to Google-Extended and Applebot-Extended, "amongst others," that they not crawl the site for AI training.⁴ That "amongst others" is as far as the documentation goes. Lyra couldn't find a published list of every user agent in the managed block. Our own served file on August 20 named six, GPTBot and ClaudeBot among them, while the file on our server said AI crawlers were welcome. So on our zone, at least, the block ran longer than the two names Cloudflare gives.¹

One more thing the documentation is plain about: robots.txt is a request. Cloudflare's own page says compliance is voluntary and that the file does not stop a crawler at a technical level.² So the managed file, on its own, changes what you're asking for. It doesn't change what's possible.

The setting that does block

That's the second feature. Cloudflare calls it AI Crawl Control, and its docs describe it as the way to enforce crawl blocking rather than request it.³ It blocks at the network edge, and nothing your robots.txt says reaches it. Our own record puts it about as plainly as it can be put: the origin said welcome, the edge said no, and the edge wins.¹ Cloudflare suggests running the two together, robots.txt to state a preference and AI Crawl Control to enforce it.³

If you meant to allow AI crawlers, this is the one to check first. A crawler can ignore a text file. It can't ignore a block at the edge.

How to see what's actually being served

From a machine that isn't your server, request your robots.txt the way any crawler would. A browser works; so does curl. Read the whole thing, top to bottom. If there's a block near the top you didn't write, the managed file is on. Then compare that against the file on your origin. If the two differ, you've found the gap.

The network-level block won't show up in that file at all. That's the point of it. For that one you have to go into the dashboard.

Where the switches are

For the managed file, Cloudflare's instructions go to the Security Settings page, filter by Bot traffic, and find the setting labeled "Set your preference to block training in robots.txt."² That's the switch. On prepends the block. Off stops it. We didn't go through the dashboard ourselves. One PUT to the zone's bot_management endpoint turned both off, ai_bots_protection set to disabled and is_robots_txt_managed set to false. The exact request is in the evidence file.¹

About the word "silently"

The assignment that produced this piece assumed these features arrive switched on. Lyra couldn't confirm that. Cloudflare's blog says "once enabled," which reads as something a person turns on, and she found no source saying either feature is on by default on any plan.⁴ Our own case doesn't settle it. Nobody here turned them on deliberately; that much we know. How they ended up on, we don't. So I'm not going to tell you Cloudflare did this to you. What I can tell you is that it's a switch in a security menu, and its label describes a preference you could flip while thinking about something else entirely.

Whether that happened on your account is a quick check. The file on your server is your intent. The file the internet receives is your policy. Go read the second one.

This check is one of the things our AI visibility audit runs from the outside, along with the rest of what a crawler sees that you don't. If you'd rather not go through the dashboard yourself, that's the page.