Mostly different businesses. Of the 2,405 business names Gemini Flash-Lite gave from memory, only 20.3% appeared anywhere in Claude's live-search answers for the same industry: 25.0% in English and 14.8% in Arabic. Memory is a short list of established names. In home services, the five names Gemini mentioned most took 46% of all its mentions; in live search, the top five took 10%.
For a small business, that is the useful part. Live search answers each question from the pages that fit it, so 81% to 98% of the businesses it named in each industry came up in just one answer. A specific page can get you named soon. Memory favours businesses that have been written about for a long time, and it only changes when a new model is trained.
What do "memory" and "live search" mean?
An AI assistant can answer a "who should I use?" question in two ways. From memory, it writes the answer from what the model learned in training, with no new lookup. With live search, it runs a web search first and builds the answer from the pages it finds, usually with links to them. Which one happens depends on the tool, its settings and the question.
Our study measured each one on its own. Gemini Flash-Lite answered 478 questions through the Gemini API with no search at all (Google Search grounding is not part of the API's free tier). Claude answered all 500 with Anthropic's live web search tool. Both ran on 21 October 2026, with the same questions word for word and the same rules for coding every business named. The method is in the main study.
One caution. The two lenses are also two different companies' models, so part of the gap is Gemini against Claude, on top of memory against search. A small check suggests model size matters too. On 10 English dental questions, Gemini 3.8 Flash, a bigger Gemini model also answering from memory, had 14 of its 53 names (26%) in Claude's live-search dental answers. Gemini Flash-Lite, on the same 10 questions, had 5 of 46 (11%). Ten questions show a direction only.
How many remembered businesses show up in live search?
About one in five. For each name Gemini gave from memory, we checked whether it appeared anywhere in Claude's 50 live-search answers for the same industry, in any question.
| Industry | Gemini's remembered names found in live search |
|---|---|
| Real estate | 50.7% |
| Business setup | 46.6% |
| Home services | 24.8% |
| Education and childcare | 21.9% |
| Law firms | 18.0% |
| Fitness and wellness | 16.9% |
| Aesthetics and dermatology | 16.8% |
| Dental clinics | 8.5% |
| Restaurants and cafes | 7.3% |
| Automotive | 6.1% |
| All industries | 20.3% (English 25.0%, Arabic 14.8%) |
The overlap is highest where a few big brands lead both the category and the web, like property brokers and company setup consultants. It is lowest where the market is local and crowded: garages, restaurants and dental clinics. There, what the model remembers and what the web shows today are almost separate lists. In Arabic the overlap drops again, partly because names are spelled differently in Arabic and simple matching misses some pairs.
The result holds however you count. Even with very generous matching (partial names, plus the Latin spelling in brackets, which over-counts), the share only reached 33.0%. The other direction is just as wide: of the 1,975 business names in Claude's live-search answers, 11.7% matched a name Gemini gave from memory for the same industry.
Why does memory keep naming the same few businesses?
Because a model remembers the names that were written about most. The table shows how concentrated each lens is. The top-5 share is the part of all mentions in an industry taken by its five most-named businesses. "Named in one answer only" is the share of businesses that appeared in just one of the 50 answers.
| Industry | Top 5 share, memory | Top 5 share, live search | Named in one answer only, memory | Named in one answer only, live search |
|---|---|---|---|---|
| Home services | 46% | 10% | 77% | 82% |
| Business setup | 39% | 13% | 73% | 81% |
| Real estate | 37% | 13% | 77% | 83% |
| Law firms | 26% | 14% | 75% | 85% |
| Aesthetics and dermatology | 15% | 7% | 79% | 91% |
| Automotive | 15% | 7% | 86% | 93% |
| Fitness and wellness | 12% | 6% | 89% | 91% |
| Dental clinics | 10% | 10% | 85% | 86% |
| Education and childcare | 10% | 7% | 81% | 90% |
| Restaurants and cafes | 6% | 4% | 95% | 98% |
The top-5 shares are the study's figures. The last two columns are our count from the raw answers with the same name matching, so spelling variants of one business may count twice. In business setup, 27% of the businesses Gemini named from memory came back in two or more answers, against 19% in live search. In home services, three names carried the memory answers.
The big names, side by side
These are the businesses Gemini mentioned most from memory, and how many live-search answers mentioned them, counting any spelling in Arabic or Latin script. This is what two AI assistants said on one day, 21 October 2026. It is not our ranking and not a judgement of any business.
| Industry | Business, as the AI named it | Gemini from memory (answers) | Claude live search (answers) |
|---|---|---|---|
| Home services | ServiceMarket | 32 of 44 | 0 of 50 |
| Home services | Urban Company | 27 of 44 | 6 of 50 |
| Home services | Justlife | 23 of 44 | 4 of 50 |
| Real estate | haus & haus | 27 of 50 | 2 of 50 |
| Real estate | Betterhomes | 23 of 50 | 9 of 50 |
| Real estate | Allsopp & Allsopp | 22 of 50 | 4 of 50 |
| Real estate | Driven Properties | 18 of 50 | 6 of 50 |
| Business setup | Shuraa | 25 of 37 | 16 of 50 |
| Business setup | Decisive Zone | 20 of 37 | 0 of 50 |
| Business setup | Virtuzone | 15 of 37 | 3 of 50 |
| Law firms | Al Tamimi & Company | 25 of 50 | 4 of 50 |
| Law firms | Davidson & Co | 14 of 50 | 0 of 50 |
| Fitness and wellness | Fitness First | 18 of 50 | 8 of 50 |
| Fitness and wellness | GymNation | 7 of 50 | 6 of 50 |
| Aesthetics and dermatology | Novomed | 11 of 50 | 4 of 50 |
| Aesthetics and dermatology | Kaya Skin Clinic | 9 of 50 | 5 of 50 |
| Automotive | Royal Swiss Auto Services | 9 of 50 | 0 of 50 |
| Automotive | Orange Auto | 8 of 50 | 0 of 50 |
| Dental clinics | Micris Dental Clinic | 5 of 50 | 0 of 50 |
| Restaurants and cafes | Calicut Notebook | 7 of 48 | 2 of 50 |
| Restaurants and cafes | Al Mallah | 6 of 48 | 1 of 50 |
Some of the names memory leaned on hardest did not come up in live search at all. Others held up in both, like Shuraa, Fitness First and GymNation, and those had their own pages in the live answers' sources: Shuraa's site was cited 24 times in business setup answers and GymNation's 16 times in fitness answers. Being well known got them into memory. Having pages that answer the question kept them in live search.
What does this mean for an established business?
If you are a well-known name, memory works for you. When an assistant answers without searching, you are likely on its list, often near the top. But live search does not reward fame on its own. It answers from the page that fits the question, and in live search the 50 real estate answers were spread over about 145 different names. A big brand with one generic page per service loses the specific questions ("Best broker for renting a 2 bedroom apartment in Ajman?", "Which agency in Ajman charges the lowest commission?") to whoever wrote that page.
Memory also ages. A model only knows what was written before it was trained, so a new branch, a new service or a new price is invisible to memory until a later model learns it. Keep your facts the same everywhere, and build pages for the specific questions your buyers ask, so you are covered both ways.
What does this mean for a small business?
Live search is where you can win, and win quickly. In live search, the five most-named businesses took only 4% to 14% of mentions in each industry, and most named businesses came up once. Many were small local firms.
AC repair near Al Barsha today?
Claude with live search named seven local firms, starting with AtoZ Appliances Repair, Jabal Albaz Technical Services, Dubai Cool Experts, UAE Quick and Home Mate Dubai. Gemini from memory named Justlife, Mr. Usta and Urban Company.
Memory went to the platforms everyone has heard of. Live search went to local firms with pages about AC repair in that area. For a small business, the job is to be the page that answers the exact question, with the area, the service, the price and how fast you can come. The way into memory is slower: be described consistently in many places (directories, reviews, local press, lists) so that future models learn the same facts about you.
Which kind of answer will my buyer see?
You cannot choose. It depends on the assistant, its settings and the question, and it can change from one question to the next. A practical sign: an answer with links to sources searched; an answer with no links probably came from memory. So when you check what AI says about your business, check both kinds of answers, and count how often you are named as well as whether you come first.
What we'd do for you
In an AI visibility audit (AED 1,950) we ask your buyers' questions, 30 per language, on four AI tools and show which answers name you, which name your competitors, and which websites those answers rely on. On a monthly plan we build the pages live search needs for each question, work on the outside mentions that shape what models remember, and re-check the same questions on ChatGPT and Gemini (Google AI) on a schedule. The AI check is free.
What this data does not show
- One run of each question on one day, 21 October 2026.
- Memory and search were measured on different models (Gemini Flash-Lite and Claude), so the gap mixes both effects. The 3.8 Flash check covers 10 questions only.
- Claude's search runs on a US-based index, and Gemini Flash-Lite is Google's smaller, faster model. Neither is identical to what a person sees in the ChatGPT or Gemini apps.
- Names were matched with simple rules (lowercase, common words like "clinic" or "LLC" removed). Spelling variants and Arabic versus Latin names mean some matches are missed.
- We did not check whether every named business exists or is still open.