AI almost always names businesses: 92% of Claude's live-search answers and 94% of Gemini's answers named at least one, usually four or five. But the lists barely agree. On the same question, the two assistants shared even one business only 16% of the time, and two versions of Gemini picked the same first name in just 4 of 19 questions.
Arabic is where it gets thin. Arabic answers named fewer businesses, hedged twice as often, and at least 11 of 250 Arabic answers drifted to businesses in Saudi Arabia. For a UAE business, being named is winnable, but nobody holds the answer for long, and one check on one tool proves nothing.
Why we ran this
73.3% of working-age people in the UAE use AI tools, the highest share in the world, against 18.8% worldwide (Microsoft AI Diffusion Report, Q2 2026). A growing share of "who should I use?" questions now end in an AI answer with a short list of names. Most UAE businesses have never checked whether they are on that list, and most advice about getting there is guesswork. We wanted numbers: how often AI names a business at all, how many it names, where the answer comes from, and how stable it is.
How we did it
- 500 questions. 10 industries, 50 questions each: 25 in English and 25 in Arabic, written the way UAE buyers ask (comparing, price, urgent, trust and very specific situations), across Dubai, Abu Dhabi, Sharjah and Ajman. Examples: "Emergency dentist open now near JLT?", "أفضل شركة عقارات في دبي مارينا لشراء شقة؟".
- Lens 1: Claude with live web search. Every question was sent once, word for word, to Anthropic's web search tool on 21 October 2026. We saved the answer and every source it cited (4,634 citations in total). The search runs from a US-based index, which matters for the location findings below.
- Lens 2: Gemini from memory. The same questions went to Google's Gemini Flash-Lite through the Gemini API, with no live search, on the same day. 478 of 500 were answered before the free tier's daily limit; the other 22 are marked in the data.
- Consistency checks. 20 questions (the first English and Arabic question in each industry) also went to a second Gemini model, Gemini 3.5 Flash-Lite, and 10 dental questions to Gemini 3.8 Flash.
- Coding. Every answer was read in full and every specific business it recommended was listed, with one fixed rule set for both lenses: platforms and directories (Bayut, Talabat, Doctify and the like), government bodies, product brands and named individuals were left out. An answer that named no one, or only pointed to maps and reviews, was coded as hedged.
The question list and the per-question data are free to download in the sources below.
Finding 1: AI names businesses, almost every time
This is the first thing to understand. AI rarely refuses to recommend. Claude named at least one business in 92.0% of its answers (95.2% in English, 88.8% in Arabic), with 4.3 businesses on average. Gemini named at least one in 93.9% of answers, with 5.4 on average.
So the question for a business is not whether AI will recommend someone in its category. It will. The question is whether it will be one of those four or five names.
| Industry | Claude: named anyone | Claude: names per answer | Gemini: named anyone | Gemini: names per answer |
|---|---|---|---|---|
| Restaurants and cafes | 100% | 5.7 | 100% | 7.9 |
| Fitness and wellness | 100% | 4.1 | 100% | 4.7 |
| Dental clinics | 98% | 3.9 | 92% | 4.7 |
| Aesthetics and dermatology | 98% | 3.6 | 94% | 4.7 |
| Business setup | 96% | 3.9 | 97% | 5.3 |
| Education and childcare | 92% | 4.7 | 100% | 5.0 |
| Home services | 90% | 4.6 | 100% | 4.7 |
| Automotive | 86% | 4.4 | 94% | 5.6 |
| Real estate | 82% | 4.6 | 82% | 5.6 |
| Law firms | 78% | 3.4 | 82% | 5.4 |
Law firms and real estate are the exceptions. Both assistants hedged on about one in five of those questions (Claude 22% for law, 20% for real estate; Gemini 18% for both), often naming individual lawyers or agents, or pointing to the regulator and the portals instead of a firm. In regulated, high-stakes categories, AI is more careful, and a firm needs stronger outside proof to be named.
Finding 2: The lists barely agree
For the 421 questions both assistants answered with at least one name, Claude and Gemini shared even one business on only 16.4% of them: 25.6% in English and 6.8% in Arabic. Most of the time, the two assistants recommended completely different businesses for the same question.
Inside Google's own family it is not much steadier. On 19 questions answered by two Gemini Flash-Lite models, they shared at least one name on 14, but put the same business first on only 4. On 10 dental questions, Gemini 3.8 Flash and Gemini Flash-Lite agreed on the first name once.
Best car service garage in Dubai Marina?
First name from Gemini Flash-Lite: Deutsche Technik. First name from Gemini 3.5 Flash-Lite: Royal Swiss Auto Services.
What this means: a screenshot of one assistant naming you is a snapshot, not a position. Buyers use different assistants and different versions, and each one builds its list differently. Tracking has to cover more than one tool, on a schedule, with the same questions each time.
Finding 3: AI memory is an old list of big names; live search is wide open
We compared what Gemini names from memory with what Claude found with live search. Of the 2,405 business names Gemini gave, only 20.3% showed up anywhere in Claude's live-search answers for the same industry (25.0% in English, 14.8% in Arabic).
The shape of the two lists is different too. From memory, Gemini leans on a handful of established names. With live search, the names scatter.
| Industry | Share of all mentions taken by the 5 most-named businesses: Gemini (memory) | Same, Claude (live search) |
|---|---|---|
| Home services | 46% | 10% |
| Business setup | 39% | 13% |
| Real estate | 37% | 13% |
| Law firms | 26% | 14% |
| Aesthetics and dermatology | 15% | 7% |
| Automotive | 15% | 7% |
| Fitness and wellness | 12% | 6% |
| Dental clinics | 10% | 10% |
| Education and childcare | 10% | 7% |
| Restaurants and cafes | 6% | 4% |
In live search, 81% to 98% of the businesses named in each industry were named only once across all 50 questions. The most-named businesses were the ones you would expect from memory: in real estate, haus & haus, Allsopp & Allsopp, Betterhomes and Driven Properties; in business setup, Virtuzone and Shuraa; in law, Al Tamimi & Company; in home services, ServiceMarket, Urban Company and Justlife.
The opening for smaller businesses. When an assistant searches, it answers each question from whatever pages fit that exact question. That is why so many names appear once: a clinic, garage or nursery with a page that answers a specific question can be named for it, even without the brand size the model remembers. When it answers from memory, the established names win. Being found by live search is the part a business can change quickly.
Finding 4: Where the live answers come from
Claude cited 9.3 sources per answer on average. Here is what they were:
| Type of source | Share of 4,634 citations |
|---|---|
| Businesses' own websites and other sites | 78.2% |
| Directories and listings (Bayut, Property Finder, Practo, Fresha, TimeOut and similar) | 9.4% |
| Wikipedia | 5.8% |
| Review platforms (TripAdvisor and similar) | 4.1% |
| Social media and forums | 1.3% |
| News and media | 1.0% |
| Government sites | 0.2% |
The single most-cited site was Wikipedia (267 citations), but not for the businesses themselves: the pages were a scattered mix of companies, landmarks and organisations, some only loosely related to the question. Then came Bayut (143), TripAdvisor (106) and Property Finder (98).
38% of all citations had a "best", "top" or "أفضل" style title. In English it was 43%. Ranked lists, many written by businesses or agencies about their own category, are a large part of what the answers are built from. Restaurants were the extreme: 65% list-style sources and 27% review platforms.
Our reading. A business's own site is the biggest single input, but only for pages that answer the question asked. Directory profiles and "best of" lists fill the rest. Government pages almost never appear, even for questions about licences or rules, so a business that explains the rules clearly on its own site fills a gap.
Finding 5: Arabic is thinner, and sometimes in the wrong country
Asked in Arabic, both assistants did reply in Arabic (96.8% of Claude's answers and all of Gemini's). They also used Arabic sources: 90.2% of the citations in Claude's Arabic answers were Arabic pages. But the answers had less in them:
| English | Arabic | |
|---|---|---|
| Claude: answers naming at least one business | 95.2% | 88.8% |
| Claude: businesses per answer | 4.9 | 3.7 |
| Claude: hedged (named no one) | 5.2% | 11.6% |
| Gemini: businesses per answer | 5.8 | 4.9 |
| Claude and Gemini named a business in common | 25.6% | 6.8% |
At least 11 of the 250 Arabic answers from the live-search lens described businesses in Saudi Arabia, in Riyadh or Jeddah, instead of the UAE. It happened most when the area name also exists there (Al Nahda, النهدة, is a district in Dubai, in Sharjah and in Riyadh) or when the Arabic name of a Dubai area reads like a general phrase (Business Bay is الخليج التجاري). A few English answers drifted to the United States in the same way.
What this means in Arabic. Fewer UAE businesses have Arabic pages, so the assistant has less to work with, names fewer of them, and sometimes fills the gap with businesses in another country. When good Arabic pages exist, they get used. Write the area and the city in Arabic and in English on your Arabic pages ("النهدة، دبي"), so an assistant cannot mistake where you are.
What a UAE business should do with this
- Check more than one assistant. The lists differ so much that one tool tells you little. Ask the same 10 buyer questions on at least two assistants, in English and Arabic, and keep the screenshots with the date.
- Build a page per real question. 78% of live-search citations were websites, and answers were assembled question by question. A page that answers "emergency dentist open now near JLT" plainly, with hours, area and price, is what gets used.
- Get onto the lists and directories in your category. List-style pages and directory profiles are a large share of what assistants read, especially for "best" questions.
- Write real Arabic pages. Arabic answers are thinner and less consistent, and they use Arabic pages when they exist. Name your area and city in both scripts.
- Expect memory to lag. Without live search, AI falls back on established names. Consistent details everywhere (your site, Google Business Profile, directories, social profiles) and fresh mentions are how a business gets into what the models learn next.
- In law and real estate, build outside proof. AI hedges most in these categories. Named recognition, reviews and clear regulatory details help an assistant feel safe naming a firm.
Limits of this study
- One day (21 October 2026) and one run per question. AI answers change from day to day.
- Claude's search runs on a US-based index, so it may see different pages than a user searching from Dubai. Gemini answered without live search. Neither lens is identical to what a person sees in the ChatGPT or Gemini apps.
- Gemini Flash-Lite is Google's smaller, faster model. 22 Gemini answers are missing because of the free tier's daily limit.
- The consistency checks are small (19 and 10 questions) and show direction, not a precise rate.
- Business names were matched across answers with simple rules (lowercase, common words like "clinic" or "LLC" removed). Spelling variants and Arabic versus Latin names mean some matches are missed, so the overlap figures are lower bounds.
- We wrote the questions. We did not check whether every named business exists or is still open.