AI almost always names businesses for US buyers: 94% of Claude's live-search answers named at least one, five on average. But the answers are built from directories and lead-generation lists far more than from businesses' own websites, and one source in six was a spam-style page on a site that had nothing to do with the question. The lists are wide open: 92% of the 2,159 businesses named appeared in only one answer.
Spanish is where it gets thin. Spanish questions got Spanish answers, but only a quarter of their sources were Spanish pages, and 4 of 100 Spanish answers were built mostly on another country, such as an Argentine insurer called Boston Seguros for "cheap car insurance in Boston". For a US business, being named is winnable with a real page per service and city, and the right directory profiles.
Why we ran this
About half of US adults (49%) now use AI chatbots, and 42% use them to look up information, according to Pew Research Center (June 2026). Microsoft puts US use of generative AI at 33.0% of working-age people, against 73.3% in the UAE. More and more "who should I use?" questions end in an AI answer with a short list of names.
In October 2026 we published the same study for the UAE. We also work with US businesses, so we ran it again here with the same method, the same assistant and the same kind of questions, and the two markets can be compared line by line.
How we did it
- 500 questions. 10 industries, 50 questions each: 40 in English and 10 in Spanish, written the way US buyers ask (comparing, price, urgent, trust and very specific situations), across 20 large US cities: New York, Los Angeles, Chicago, Houston, Phoenix, Philadelphia, San Antonio, San Diego, Dallas, Austin, Miami, Atlanta, Seattle, Denver, Boston, Nashville, Charlotte, Las Vegas, Orlando and Tampa. Examples: "Emergency plumber in Charlotte?", "Dentist in Miami that takes Delta Dental and is well reviewed?", "Mejor abogado de inmigración en Orlando".
- One lens: Claude with live web search. Every question was sent once, word for word, to Anthropic's web search tool, the same tool as our UAE study, on 21 October 2026 between 21:31 and 23:12 UTC (afternoon and evening in the US). The query was the question text exactly, with nothing added. We saved the answer and every source listed with it: 4,587 sources from 1,474 different websites. All 500 calls worked.
- Coding the names. Every specific business the answer named or recommended was listed. Platforms and directories (Zocdoc, Thumbtack, Avvo, OpenTable and the like), government bodies and generic categories were left out. One difference from the UAE run: named professionals were kept, because many US lawyers, doctors and agents trade under their own name. We also computed every headline number without them, and the UAE comparison shows both. An answer that named no one, or only pointed to maps, reviews or directories, was coded as hedged.
- Coding the sources. Every source was sorted into one of nine types (see Finding 2). We built the US lists by hand from the 1,474 sites, checking titles and addresses.
- Hand check. We read 20 answers drawn at random in full and compared them with the coded names. Of 107 names, none were missed and no platform slipped in. Three were double counts (a physical therapist listed next to the clinic they work at), an error rate of about 3%. The hedged label was right on all 20.
A second lens, Google's Gemini, is being collected separately and is not part of these numbers. We will add the comparison here if and when it is complete.
Finding 1: AI names businesses, almost every time
Claude named at least one business in 93.8% of its answers (94.5% in English, 91.0% in Spanish), with 5.2 businesses on average when it named any. It hedged on only 6.2% of questions. So for most US categories the question is not whether AI will recommend someone. It will. The question is whether you are one of those five names.
| Industry | Named anyone | Names per answer | Hedged | Named once only |
|---|---|---|---|---|
| Restaurants and cafes | 100% | 7.7 | 0% | 96% |
| Fitness and wellness | 100% | 5.0 | 0% | 96% |
| Home services | 96% | 5.0 | 4% | 91% |
| Automotive | 96% | 4.6 | 4% | 93% |
| Accountants and tax | 94% | 5.5 | 6% | 79% |
| Real estate | 92% | 6.1 | 8% | 92% |
| Education and childcare | 92% | 5.0 | 8% | 95% |
| Dentists | 92% | 3.6 | 8% | 91% |
| Law firms | 90% | 5.0 | 10% | 94% |
| Med spas and dermatology | 86% | 4.1 | 14% | 93% |
Three patterns stand out. Price questions are the weak spot: on the 51 English questions about cost, only 82.4% named anyone and 17.6% hedged, against 2.4% for questions that ask for the best option. The answers kept saying they found no reliable local price. Health, beauty and law are careful: med spas hedged most (14%), then law firms (10%), and dentist answers named the fewest businesses (3.6). Law and real estate name people: a third of the names in those two industries (33%) were individual lawyers and agents, not firms. Counting firms only, 84% of law answers and 86% of real estate answers named a firm.
What is the best dentist in New York?
There's no single objective "best" dentist in New York. The results are mostly promotional pages and award announcements, so treat them as leads rather than rankings.
That caution is typical. In at least 55% of answers Claude warned that its sources were marketing, sponsored or low quality, and in at least 55% it said there was no single best option. It still named businesses in nearly all of them.
Finding 2: The answers are built from directories, not business websites
Claude listed 9.2 sources per answer. Here is what they were:
| Type of source | Share of 4,587 sources |
|---|---|
| Directories, marketplaces and "best of" list sites (Thumbtack, Avvo, Zocdoc, Niche, OpenTable, HomeLight and similar) | 37.3% |
| Businesses' own websites and other sites | 29.0% |
| Off-topic hosts (spam-style pages on unrelated sites) | 15.9% |
| News and media (city magazines, local TV, food and travel guides) | 11.0% |
| Review platforms (Google business pages, Trustpilot, Experience.com and similar) | 3.2% |
| Social media and forums | 1.6% |
| Press releases | 1.2% |
| Government sites | 0.9% |
| Wikipedia | 0.1% |
81% of answers cited at least one directory. The most-cited sites overall were Lantern, a local business directory (86 citations), Thumbtack (64), LawInfo (59), BiggerPockets (53, mostly its agent finder and forum), Expertise.com (49), Avvo (47) and HomeLight (45). Each industry has its own set:
- Law firms: 56% of sources were directories. LawInfo, Avvo, Abogado.com, Expertise.com, USLegal and FindLaw led.
- Education and childcare: 46% directories, led by Niche, Winnie, Varsity Tutors and Wyzant.
- Fitness: 43% directories, led by Lantern, Magicpin, ClassPass and Fresha.
- Real estate: 40% directories, led by HomeLight, BiggerPockets, Simplifyem and Redfin's agent pages.
- Home services: Thumbtack and HomeAdvisor alone were 61 citations.
- Restaurants: the one industry led by media. 43% of sources were news, food and city guides, and 53% of sources had a "best" or "top" style title.
- Accountants: the one industry led by business websites (49%). Two firms, Uncle Kam and KDA Inc., were cited 85 times between them.
Some big names are missing. In this lens, Yelp was never cited, yet 100 answers mentioned it, most of them telling the buyer to check Yelp reviews. 135 answers pointed to Google reviews or Google Maps, but only 18 cited a Google business page. The Better Business Bureau was mentioned in 21 answers and cited in none. Tripadvisor, Facebook, Instagram, YouTube and Nextdoor did not appear as sources at all, and Reddit appeared only once, through a mirror site. Wikipedia, the most-cited site in our UAE study, was cited 5 times.
Our reading. The search behind these answers favours pages whose titles match the words of the question. Directory pages titled "Top Denver divorce lawyers" or "Best daycares in Dallas" match those words exactly, so they make up the backbone of the answer. Review sites still matter, because the answer sends the buyer there next, but in this lens your directory profiles are what the assistant actually reads.
Finding 3: One source in six was a spam-style page on an unrelated site
This was the surprise. 731 sources, 15.9% of the total, were pages about a local business topic sitting on a website that has nothing to do with it. They came from 303 different sites and showed up in 53% of answers. In 47 answers they were most of the sources.
We call them off-topic hosts. Typical examples, described without naming the sites, which are often the victims here:
- a university server in Portugal carrying a set of "lip filler cost" and "DUI lawyer cost" guides (11 pages, cited 25 times)
- a photography tutorial site with "best dentist in San Diego" pages
- a US state government subdomain with a list of "top-rated gyms in San Diego"
- a hardware supplier's subdomain with CoolSculpting tips, and a cancer charity's site with a pizza guide
- forum profiles, free page builders, staging servers and code wikis stuffed with "best X in city" text
Many are mass-produced: the same titles appear on several unrelated hosts, and some still contain template text like "[Shop Name 1]". Our count is a floor, because we only flagged hosts we confirmed by hand.
The answers often noticed, calling such pages "SEO filler", "low-quality" or "unreliable". But noticing is not ignoring: a page titled "best orthodontist in Miami" still shapes which names appear.
The UAE answers did not have this problem. The same automatic checks flagged 16 of 4,634 UAE sources; half were real government sites in Saudi Arabia and Bahrain, and most of the rest were ordinary business pages. In this lens, the spam problem is a US problem.
What this means for you. Your competitor for a spot in the answer is sometimes a spam page that happens to match the question's words. A real page on your own site that answers the same question plainly (service, city and state, price, hours) competes with it directly, and does not vanish when a hacked server is cleaned up. Do not pay for placement on pages like these; they can disappear overnight, and the answers already treat them with suspicion.
Finding 4: The lists are wide open
Across the 10 industries, the answers named 2,159 different businesses in 2,425 mentions. 92.5% of them were named only once in 500 answers. The five most-named businesses in an industry took between 3% (restaurants) and 15% (dentists) of all mentions. Nobody owns these answers.
| Industry | Named most often in our answers (number of answers) | Top 5 share of mentions |
|---|---|---|
| Dentists | Aspen Dental (7), Absolute Dental (7), Affordable Dentures & Implants (6) | 15% |
| Accountants and tax | Uncle Kam (9), Deloitte (5), KDA Inc. (5), 1-800Accountant (5) | 11% |
| Automotive | Tint World (8), GEICO (5), State Farm (4), Progressive (4) | 11% |
| Home services | Roto-Rooter (6), Orkin (5), One Hour Heating & Air Conditioning (3) | 8% |
| Real estate | Redfin (7), Compass (4), RE/MAX (3), plus individual agents | 8% |
| Med spas and dermatology | Cultura Dermatology (3), Dallas Center for Dermatology and Aesthetics (3), Ideal Image (3) | 7% |
| Fitness and wellness | Planet Fitness (5), Equinox (4), 24 Hour Fitness (3) | 7% |
| Education and childcare | General Assembly (4), CodeWizardsHQ (3), Bright Horizons (2) | 6% |
| Law firms | Morgan & Morgan (3), Vásquez Law (3), Baker Law Group (2) | 5% |
| Restaurants and cafes | Factor (3), Taquiza (3), with no other name above 2 | 3% |
These are counts from our 500 answers, not rankings or recommendations.
The names that do repeat tend to share a habit: many local pages. Uncle Kam, a tax firm, was named in 9 accounting answers, and its pages, down to ZIP code level, were cited in 11 answers across 7 cities. Affordable Dentures & Implants was cited in 6 dental answers in 6 cities, through its location pages for each office. Tint World, a franchise with local stores, was cited in 8 auto answers in 8 different cities. National chains and insurers repeat for a similar reason: most have a page for nearly every city.
The opening for smaller businesses. When the assistant searches, it answers each question from whatever pages fit that exact question. That is why 92% of names appear once: a clinic, garage or tutor with a page that answers one specific question well can be named for it, without a national brand. Being found by live search is the part a business can change quickly.
Finding 5: Spanish answers lean on English pages, and sometimes on other countries
99 of the 100 Spanish questions got an answer in Spanish. The hundredth, a question about skin clinics in San Antonio, came back in Russian, describing four clinics in Mexico. That one is an outlier, but it points at the real issue: the sources.
| English | Spanish | |
|---|---|---|
| Answers naming at least one business | 94.5% | 91.0% |
| Businesses per answer | 5.3 | 4.6 |
| Hedged (named no one) | 5.5% | 9.0% |
| Sources with a "best" or "top" style title | 28.9% | 15.8% |
| Sources that were Spanish-language pages | 0.5% | 23.8% |
Only 23.8% of the sources behind Spanish answers were Spanish pages, and 41% of Spanish answers cited no Spanish page at all. In our UAE study, 90.2% of the sources behind Arabic answers were Arabic pages. The difference is supply: few US businesses publish real Spanish pages, so the assistant reads English pages and translates. Where Spanish pages existed, they were used heavily. In law, 9 of the 10 Spanish answers cited Abogado.com (31 citations in all) or one firm's Spanish blog (14).
When Spanish pages are missing, the search sometimes finds them in another country. We read every Spanish answer that mentioned a foreign place:
- 4 of 100 were built mostly on another country: whitening prices from Spain and Chile for Nashville, four clinics in Mexico for San Antonio, lip filler prices from Spain and the UK for Charlotte, and car insurance from an Argentine company named Boston Seguros for Boston.
- 7 more had sources from abroad that the answer set aside or did not use: San Antonio in Chile, Los Ángeles in Chile, a Los Ángeles district in Murcia, Spain, driving schools in Durango, Mexico, clinic pages in Spain and Australia, and Mexican yellow pages.
English answers did not drift abroad. Their foreign websites were mostly spam hosts or foreign versions of US directories (OpenTable's Mexican and Spanish sites, for example) writing about the US city asked. A handful were about the wrong country, such as a London leisure centre named Phoenix for a Phoenix gym question, and no English answer used them.
Seguro de auto barato en Boston
Los resultados no sirven para Boston, Massachusetts. Casi todos hablan de Boston Seguros, una aseguradora argentina, con páginas para localidades como Urdinarrain, Allen, Ramos Mejía y Daireaux.
Same-name places trip up English answers too. In at least 11 answers, sources were about a different place with the same name: Denver, Iowa and Denver, North Carolina; Dallas, Oregon, Georgia and Wisconsin; Chicago, Indiana; Nashville, North Carolina. Usually the answer caught it; once, a business from the wrong Denver still made the list.
What this means in Spanish. Spanish answers are thinner because Spanish pages are scarce, so a business with a real Spanish page per service has little competition. Write the city and the state on every page, in both languages ("Boston, Massachusetts", "San Antonio, Texas"), so an assistant cannot mistake where you are.
UAE and US, side by side
Same assistant, same search tool, same design: 500 questions, 10 industries, two languages. The industry mix differs slightly: the US study has accountants and med spas where the UAE study had business setup and aesthetic clinics.
| Claude with live web search | UAE (Oct 2026) | US (Oct 2026) |
|---|---|---|
| Answers naming at least one business | 92.0% | 93.8% (92.2% counting firms only) |
| Businesses per answer, when any | 4.3 | 5.2 (4.7 firms only) |
| Hedged | 8.4% | 6.2% |
| Sources per answer | 9.3 | 9.2 |
| Business and other websites | 79.4% | 44.9% (of which 15.9% off-topic hosts) |
| Directories and lists | 9.5% | 37.3% |
| Review platforms | 2.9% | 3.2% |
| News and media | 0.9% | 12.2% (incl. 1.2% press releases) |
| Social media and forums | 1.3% | 1.6% |
| Wikipedia | 5.8% | 0.1% |
| Government sites | 0.2% | 0.9% |
| Sources with a "best" or "top" style title | 37.7% | 26.3% |
| Second language: named anyone | 88.8% (Arabic) | 91.0% (Spanish) |
| Second language: sources in that language | 90.2% | 23.8% |
How to read it. The UAE figures are as published. The UAE classifier had shorter lists and no off-topic or press release type, so to be fair we also re-coded the UAE sources with the US lists: business sites fell to 75.1% and directories rose to 12.7%. The gap is real. UAE answers lean on businesses' own sites and, in Arabic, on Wikipedia. US answers lean on directories, lead-generation lists and city media, and carry a spam layer the UAE answers do not.
What is the same in both: AI names someone nearly every time, the names scatter (the top five took at most 15% of an industry's mentions in either study), and the second language gets less to work with.
By city
Each city had 20 to 31 questions, too few to rank cities. Answers naming anyone ran from 87% (Nashville) to 100% (Las Vegas and Los Angeles), and off-topic hosts from 6% of sources (Seattle) to 28% (New York), which mostly reflects which industries landed in each city.
What a US business should do with this
- Fix the directory profiles that show up for your category. They are the largest single input. For lawyers that means Avvo, LawInfo, FindLaw and, for Spanish, Abogado.com; for dentists and clinics, Zocdoc, WebMD and RealSelf; for home services, Thumbtack, HomeAdvisor and Angi; for schools and childcare, Niche, Winnie and GreatSchools; for gyms and studios, ClassPass and Fresha; for agents, HomeLight and the brokerage profile pages. Same name, address, phone, services and prices everywhere.
- Publish one page per service and city. The businesses named most often across cities had exactly that. "Emergency plumber in Charlotte, North Carolina", with hours, area, price range and how fast you arrive, is a page an assistant can use.
- Put prices on the page. Price questions are where AI hedges most (17.6%) because it cannot find local numbers. A clear "from" price or a typical range gives it something to quote.
- Write city and state, every time. Same-name places sent answers to Denver, Iowa and to Chile. "Denver, Colorado" on the page costs nothing.
- Name your people, and link them to the firm. In law and real estate a third of the names AI gave were individuals. Each lawyer or agent page should say clearly which firm they work for, so a mention of the person also counts for the firm.
- If you serve Spanish speakers, write real Spanish pages. Not a translated home page: one page per service, in Spanish, with the city and state. Spanish answers have little to choose from.
- Keep Google and Yelp reviews healthy anyway. The answer sends the buyer there to check you.
- Check more than one assistant. Ask the same 10 buyer questions on ChatGPT, Gemini and Claude, keep dated screenshots, and repeat each month.
If you want help
This is the work we do for US businesses, remotely from Dubai, with traditional SEO included and ChatGPT and Gemini (Google AI) tracked on every plan. It starts with a free AI check: 10 real buyer questions, with the competitors AI names instead of you. The AI Visibility Audit (30 questions, 4 AI tools, 3 competitors, screenshots and a 90-day plan, in 10 working days) is USD 549, credited to month one if you sign within 30 days. Monthly plans are USD 1,249 (Foundation), USD 2,349 (Growth) and USD 4,399 (Authority) a month, month to month, with no lock-in and no setup fee. Full details are at adjoltz.com/us.
In the first 30 days of any plan, if AI does not name you for at least 3 commercial questions on the list you approved, or you are unhappy for any other reason, we refund 100% of what you paid, in cash, within 7 working days. The rules are on our guarantee page. What we don't promise: a top spot on any assistant, guaranteed rankings, paid places on "best of" lists, or a US office. We do not have one.
Limits of this study
- One assistant, one day (21 October 2026) and one run per question. AI answers change from day to day and between assistants. Pew found 44% of US adults use ChatGPT and 6% use Claude, so the source mix here is not what every buyer sees.
- Claude's search uses its own index. Other assistants, such as Google's, draw on different indexes and may cite Yelp, Google reviews or Reddit far more. A Gemini lens is being collected and will be compared here if it is completed.
- Spanish was 100 questions, 10 per industry, so Spanish figures by industry are directional only. City samples are small for the same reason.
- Source types were coded by us. We checked every site with two or more citations and every site that matched our spam checks by hand, but some single-citation sites may be in the wrong group. The off-topic share is a floor.
- Business names were matched with simple rules (case, punctuation, "LLC", "Inc", "Dr", "DDS" and a trailing city name removed), so the "named once" figure may be slightly high. The answer checks (warnings, "no single best") use tested text rules that miss some phrasings, so they are floors.
- We wrote the questions. We did not check whether every named business exists, is licensed or is still open, and a business being named is not an endorsement by us or by the assistant.
Downloads
Both files are free to use with credit to Adjoltz and a link to this page.
- The 500 questions (JSON lines): id, industry, language, city, intent and the exact question text.
- One row per question (CSV): industry, language, city, intent, question, number of businesses named, hedged, the first business named, number of sources, and how many sources of each of the nine types the answer used.
Our UAE study, with the same method in English and Arabic, is at UAE AI visibility study 2026.