91 names on ChatGPT, 97 on Gemini, 1 in common: in the US there is not one shortlist, there are two
We asked two AI assistants the same question, for four trades across five US cities. Both answer with confidence. They almost never name the same businesses.
1 in 188
names cited identically by both AI assistants
In this article
Short answer: in the United States, being recommended by AI is not one thing, it is two. We asked ChatGPT and Gemini the same question, for four trades across five cities. ChatGPT named 91 businesses, Gemini named 97. Exactly one appears identically in both lists. Counting the cases where the same business is written differently, we get six. And those six share something that changes the picture.
7.1
businesses named on average in one answer
0
independent practices recommended by both engines at once
Optimising for one AI assistant does not place you in the other. On this measurement, the two lists barely touch.
What we measured, and what we did not
Four trades: attorney, dentist, plumber, real estate agent. Five cities: New York, Los Angeles, Chicago, Houston, Phoenix. One question, phrased the way a resident would phrase it, plainly asking for the best professionals in town. Two AI assistants queried, ChatGPT and Gemini.
80 usable answers, and the gaps have to be stated first. We had planned a third engine, Perplexity: all 40 calls failed on an exceeded quota. This measurement says nothing about it. It is also small: our French study covers 1,386 answers, this one 80. What it supports is a gap wide enough that a sample this size makes it visible, not a fine measurement.
The two lists do not overlap
Across the 19 trade plus city pairs where both AI assistants answered, we removed entries that are not businesses, those answers where the model lists selection criteria rather than names. That leaves 91 names on the ChatGPT side and 97 on the Gemini side.
Exactly one is identical across both lists. Not one in ten, not one in five: one in 188.
Exact-name comparison is harsh, and it understates the overlap: one engine writes « Dr. Michael Apa », the other writes « Apa Aesthetic », and it is the same address. Reading every pair by hand to count the real businesses behind the labels takes us from one to six. Six out of 188.
The six exceptions tell the real story
| Business | Trade | City |
|---|---|---|
| Compass | real estate agent | New York |
| Brown Harris Stevens | real estate agent | New York |
| The Agency | real estate agent | Los Angeles |
| Josh Altman | real estate agent | Los Angeles |
| Power Rogers | attorney | Chicago |
| Rooter Hero Plumbing | plumber | Los Angeles |
Four of the six are leading brokerages or an agent known to the general public. The fifth is a Chicago law firm whose name has circulated for decades. The sixth is a plumbing franchise operating across several states.
No independent practice, in any of the five cities, is recommended by both engines at once. The neighbourhood dentist, the local plumber, the two-partner law firm: they show up on one or the other, never on both.
That is the finding, and it reads both ways. It is discouraging, because visibility work done for one AI assistant does not carry you into the other. It is encouraging, because the list you are missing from is not locked up by your competitors: it is simply a different list.
The two engines do not even recommend the same kind of thing
A deeper difference appears when you read the answers instead of counting them. Asked about dentists in New York, ChatGPT answers with a list of people: « Dr. Michael Apa », « Dr. Steven Gold », « Dr. Rebecca L. Wexler ». Gemini answers with a list of practices, with their web address: « Apa Aesthetic », followed by a link, a neighbourhood, and a sentence on what they are recommended for.
The unit of recommendation is not the same. On one you are a practitioner, on the other you are an address. That changes what there is to work on: a named biography and a professional reputation on one side, a named and described website on the other.
Three answers out of 80 name nobody at all and list criteria instead, the quality of communication, client reviews, licensing and insurance. The model judged the question too subjective to answer with names. It is rare, but it is not nil: in those cases, the question crowns no one.
An observed failure worth recognising
In three ChatGPT answers, the list falls apart after the fifth name. You read a run of initials separated by dots, then another. The model ran out of what it knew and kept producing the shape of a list without the content.
This is not a curiosity. It says that the length of an answer is not the measure of its substance, and that the low positions in a ranking produced by an AI assistant are often worth less than the high ones. On this measurement the average is 7.1 names per answer; the last two or three are regularly the most fragile.
What a business owner can do with this
Measure the two separately.
A single AI visibility score hides the one fact that matters here: you can be first on one and absent from the other.
Look at who occupies the list you are missing from.
If they are national brands, the ground is contested differently than if they are your neighbours.
Check the form you would be cited in.
A practice whose trading name resembles nothing nameable has a different problem from a well-known practitioner with no website.
Measure again.
A reading is a photograph. Models change, corpora change, a September list says nothing about December.
The limits
- 80 answers, four trades, five cities. Enough to see a massive gap, not to size it precisely, and not to speak for other trades or other cities.
- Perplexity is absent from this reading: its 40 calls failed on a quota. We do not know where it stands.
- The count of real overlaps, six, rests on a human reading of name pairs. A different reading might find five or seven, not fifty.
- An AI assistant answer depends on how the question is phrased. We asked one per trade and city; another wording would surface other names.
The question to ask is not whether AI recommends you. It is which one, and to whom. Measuring it is free and takes a minute.