Do AI Assistants Recommend US Local Businesses? In Two of Four Measured Cities the Winner Was a National Directory
New York, Los Angeles, Chicago and Houston measured on ChatGPT and Gemini on 15 and 16 October 2026: 43 of 48 readings named sources, and half the cities answered a local question with a national page.
Published by AI Knows Us · Updated 16 October 2026
They name businesses, and in half the American cities measured the business named was not local. Four metros were read on 15 and 16 October 2026, six buyer questions apiece on ChatGPT and on Gemini, for 48 readings in total. Forty three of them returned a named source list. The leader in Los Angeles and in Chicago was agencies.semrush.com, a national platform's own directory. The leader in New York and in Houston was a single city domain. Not one of the 48 reached for a page of ours.
So the answer to the question in the title has two halves in the United States, and which half you get appears to depend on your city.
City by city, and the source that led each one
Every one of the 48 readings was usable. The table below is the whole American city dataset we hold as of 16 October 2026.
| City | Readings | Readings that returned a named source list | Most cited sources in that city |
|---|---|---|---|
| New York | 12 | 9 | localseo.nyc 3, thenycseo.com 3, goodfirms.co 3 |
| Los Angeles | 12 | 11 | agencies.semrush.com 4, brightlocal.com 3, help.brightlocal.com 3 |
| Chicago | 12 | 12 | agencies.semrush.com 3, chicagoseoagency.co 3, paramdigitalmarketing.com 3 |
| Houston | 12 | 11 | houdigitalmarketing.com 3, thedigitalelevator.com 2, axios.com 2 |
| All four | 48 | 43 | 0 readings cited aiknowsus.com |
The scan read ChatGPT and Gemini. It did not touch Perplexity, Claude or Google AI Overviews, so three of the five surfaces are simply absent from every figure below.
Half of these cities answered with a page from out of state
New York and Houston went the way most local marketing advice assumes. The top sources were localseo.nyc, thenycseo.com and houdigitalmarketing.com: domains whose names are the city, run by small firms, cited because their pages are specifically about that place.
Los Angeles and Chicago did something else. The most cited source in both was a subdomain of a national software platform, used four times in Los Angeles and three in Chicago. Los Angeles also returned brightlocal.com and its help subdomain. A business in those two cities asking why it is absent from the answer is not losing to a better local competitor. It is losing to a national page that aggregates local firms.
That changes what the work is. Displacing a small city domain means publishing a better page about your city. Appearing in an answer that is being drawn from a national aggregator means getting into the thing being aggregated as well as publishing your own page, and the two jobs cost different amounts.
Where the citations go, on a much larger sample
The four city readings sit inside a wider run: 1,260 readings on 16 October 2026, which produced 20,143 citations spread over 4,947 separate domains. Sorted by what kind of site they landed on:
| Source type | Share of 20,143 citations |
|---|---|
| Vendor's or agency's own website | 82.0% |
| Official platform documentation | 12.3% |
| News or large publisher | 3.2% |
| Social or community | 1.9% |
| Review directory | 0.7% |
agencies.semrush.com belongs in the first row, not the last. It is a software company's own website, which is what 82.0% of these citations are. Review directories, the category most local businesses are advised to concentrate on, accounted for 0.7%. On this evidence a page you control is the asset, and a run eleven days earlier in design reached the same conclusion from a different question set, which is set out with its full source table in our first published benchmark of 597 readings.
The United States was the hardest of the three markets for us, twice
Two independent runs eleven days apart in design say the same thing about our own position in America. On 15 October 2026 our domain was cited zero times in the United States, on ChatGPT and on Gemini. On 16 October 2026, across 420 United States readings, it was cited once: one of 210 on ChatGPT and none of 210 on Gemini.
Across both runs together our domain was cited 3 times in 597 readings and then 6 times in 1,260. Those are two different question sets reaching the same answer, which is the only reason either of them is worth printing. A number that survives a redesigned question set is a number you can plan against.
What these 48 readings do not establish
One category was asked about, not seven. Each city's six questions came from a buyer shopping for help with AI answers, which means this scan holds no figure for a restaurant, a hotel, a gym, a salon, a contractor, a dealership or a brokerage. Each of those needs its own list written in its own words, and until somebody runs it the rate for that trade is an open question rather than an estimate.
Four cities were measured. Dallas, Phoenix, Atlanta and Miami were not, and no number above applies to them. For a city outside the four, the national picture is what you have, labelled as national.
Two surfaces of five were read. The product measures five; this scan covers ChatGPT and Gemini, and those two behaved differently enough from each other that blending them would have hidden it.
How to get your own city's answer this week
Write six questions in the words a customer in your metro types, with no company name in any of them. Run each on ChatGPT and Gemini in a fresh session, and record two things per answer: whether a source list came back at all, and which domains were in it. That is 12 rows and about an afternoon, and it tells you whether your city is a New York or a Los Angeles.
Then read the winning pages. If the top source is a single city domain, you are looking at the page you have to beat, and it is usually a small firm's own article. If the top source is a national platform's directory, your plan needs both a page of your own and a presence in the aggregate, and you should price that before you commit to it. Deciding what a single location can realistically win is the same arithmetic as what a shop in your metro charges and what proximity actually buys you.
Plain interest statement, because it changes how you should read the table: this is published by the vendor of an AI visibility product, and the strongest number on the page is the one where our own domain appeared in none of the 48 readings.
What a US local operator asks about these four cities
Which AI assistants name US local businesses?
Both of the two that were read. ChatGPT and Gemini returned a named source list in 43 of 48 city readings on 15 and 16 October 2026. Perplexity, Claude and Google AI Overviews were not part of this scan, so they are not covered by that figure.
Why is a national directory winning my city's answer?
In Los Angeles and Chicago the most cited source was a national platform's agency directory, used four and three times across 12 readings each. A page that lists many local firms can answer a local question without being local, and on this data that happened in half the cities measured.
Should we build out review directory profiles instead?
Review directories were 0.7% of 20,143 citations recorded on 16 October 2026. The 82.0% figure, which is pages on a vendor's or an agency's own site, is where the citations actually went, and that is the side of the ledger your own publishing sits on.
In how many of these readings did somebody else take the slot?
All 48. Every cited domain across the four cities belonged to another company. We publish that about ourselves because it is the baseline the next round gets compared against.
Does being cited mean being recommended?
Two different events. One is an engine leaning on a page to build its answer. The other is the answer telling a reader to go and hire that company. Keep them in separate columns when you run your own test, because a page can supply the facts while somebody else gets the suggestion.
Forty three answers went out with somebody's name in them
Not one of those 43 answers was invented on the spot. Each was assembled from pages that already existed on the day it was asked, and in two metros the page belonged to a firm down the road while in two it belonged to a software company in another state. Somebody got named either way, and in these four cities it was never us. Finding out who is filling that slot in a given metro, and publishing the page that argues for a different name, is the job our product does: AI Knows Us tracks it city by city, and its own score across New York, Los Angeles, Chicago and Houston is zero. Your metro's list is not the one in the table above, which is the whole reason to go and get it.