Twenty-one seats, and three come from your city
For a city and a trade, ChatGPT and Gemini seat a median of 21 names, and only 3 of them appear on the map of actual local businesses. The shelf is not the map; it is fame.
21
median seats per city and trade, across 4 answers (2 assistants × 2 phrasings)
In this article
Short answer: for one trade in one city, everything ChatGPT and Gemini will ever agree to name fits on a shelf of 21 seats, at the median. Lay that shelf over the map of businesses actually recorded in the city, and only 3 names are on both. The rest are national franchises, famous addresses, and a few crumbs of advice. The AI’s shelf is not the map of your city: it is the map of fame.
3
of those names, at the median, found on the map of local businesses
≤ 11.8%
of recorded businesses with a website are ever approached by an AI name
Being absent from the map is survivable. Being absent from the shelf makes you invisible.
What was measured
Two of our own datasets, crossed for the first time. On one side, the United States study (13 cities, collected 13 September 2026): every distinct name ChatGPT and Gemini returned for each city × trade pair, over 4 answers, cleaned with the rules of the previous editions. On the other, the map: businesses of those same cities and trades read from OpenStreetMap on 20 September, deduplicated by website and kept only where the city holds at least 25 - which leaves 36 comparable pairs. Name matching is deliberately generous (containment, or one shared distinctive word), so the percentages lean high, not low. Overall: 133 of the 722 AI names (18.4%) match a business on their city’s map.
| City | Trade | On the map (≥) | AI seats | Matched |
|---|---|---|---|---|
| New York | Hair salon | 171 | 19 | 3 |
| Seattle | Hair salon | 118 | 25 | 4 |
| New York | Dentist | 95 | 22 | 6 |
| New York | Attorney | 69 | 22 | 5 |
| New York | Real estate agent | 61 | 19 | 10 |
| Los Angeles | Hair salon | 55 | 21 | 0 |
| Austin | Dentist | 38 | 24 | 9 |
| Houston | Auto repair shop | 27 | 22 | 0 |
Who holds the seats: the names that travel
If the seats do not come from the city, where do they come from? From names that recur city after city. Across the 13 cities asked, the same names sit on the shelf almost everywhere:
| Trade | Names present in several cities (out of 13) |
|---|---|
| Plumbers | Roto-Rooter (10/13) · Benjamin Franklin Plumbing (8) · Mr. Rooter (4) |
| Insurance brokers | HUB International (10/13) · Brown & Brown (8) · Gallagher (8) |
| Accountants | RSM US (9/13) · Deloitte (7) · KPMG (6) · BDO (6) |
| Auto repair | Midas (8/13) · AAMCO (7) |
| Electricians | Mister Sparky (7/13) |
| Hair salons | Fox & Jane (6/13) · Spoke & Weal (6) · Butterfly Studio (5) |
| Real estate | Compass (5/13) · Keller Williams (4) · RE/MAX (3) |
A customer asking for “an accountant in Denver” is offered RSM, Deloitte and KPMG. A local plumber competes for seats with Roto-Rooter, present in ten cities out of thirteen. And among dentists, one of the best-travelled names is not a practice at all: it is Zocdoc, the platform, seated in four cities - yesterday’s mechanism, seen from the other end.
What it means if you are the business
The game is not “being a good local business”: the map records dozens, and the AI seats three. The game is getting onto a short shelf governed by fame - which can be worked: a clear entity, citations, reviews, a presence the models actually meet. That distance between existing and being findable is exactly what the audit measures.
What this does not prove
- The map is a FLOOR, not a census: OpenStreetMap does not list everything, and we kept only businesses with a website. “Absent from the map” never means “not real”: Rita Hazan and Warren Tricomi, very real New York salons named by the AI, are simply missing from our map slice.
- Generous matching over-matches: 18.4% and 11.8% are upper bounds, not fine measurements.
- The name extractor still lets advice fragments through (“consultations”, “get multiple quotes”). The visible ones were removed by hand, and every chain example quoted was checked by reading.
- United States only, 13 cities, one run per question, collected 13 September 2026. Four answers per pair: the shelf grows somewhat with more phrasings - that was the 16 September edition.
- The median of 3 matched names hides real spread: 10 in New York for real estate, 0 in Los Angeles for salons and in Houston for auto repair.
- Nothing here says why a name is seated: fame, links, reviews and training data are mixed, and this measurement does not separate them.