Do AI assistants agree on who to recommend? Not often.

What did we compare?

Three things, all on 21 October 2026, with the questions sent word for word. The method is in the main study.

The first row is the main study. The other two are small checks that show a direction, not a precise rate. Names were matched with simple rules (lowercase, common words like "clinic" or "LLC" removed), so the same business spelled two ways, or once in Arabic and once in Latin script, can be missed. Every agreement figure here is a lower bound.

How often do Claude and Gemini agree?

Seldom, and almost never on the top pick. Of the 421 questions, 352 had no business in common, 50 had one, 15 had two and 4 had three or four. Across these questions Gemini gave 2,268 names and Claude 1,829, and only 94 were the same business on the same question. Put both lists for a question together and, on average, 2.5 names in 100 appear on both (3.9 in English, 1.0 in Arabic).

Agreement was highest in education, real estate and fitness, where a few well-known schools, nurseries, brokers and gym chains lead. It was lowest for clinics, garages and law firms, where many small businesses compete area by area. In Arabic, four industries (home services, law, automotive and dental) had no business in common on a single question. Our industry reports join English and Arabic spellings of the same business, so they find a few more matches; both methods show the same picture: the two assistants rarely agree.

The 9 questions where both put the same business first

• "Best free zone in the UAE for an e-commerce business?": Meydan Free Zone
• "Which British curriculum school in Dubai is the best?": Dubai College
• "Best nursery in Abu Dhabi?": Redwood Montessori Nursery
• "Best driving school in Abu Dhabi?": Emirates Driving Company
• "Best cleaning company in Al Barsha?": Urban Company
• Best off-plan property company in Dubai for foreign investors (asked in Arabic): Emaar Properties
• Best PRO services company in Jumeirah (asked in Arabic): Shuraa Business Setup
• Best martial arts club for kids in Sharjah (asked in Arabic): Sharjah Sports Club
• Best nursery in Dubai Hills (asked in Arabic): Blossom Nursery

The pattern: a long-established name that leads its category, or a free zone, a developer or a sports club that is close to the obvious answer. Where many businesses compete, the assistants went separate ways.

Where they agreed on nothing

The first three businesses each assistant named, on questions with no name in common. This is what the AI named on one day; it is not our ranking.

Does one AI model agree with itself?

Not reliably. We sent the first English and the first Arabic question in each industry (20 questions, mostly about Dubai Marina) to Gemini Flash-Lite a second time, between 1 and 23 minutes after the first run. We set this up as a check against a second model, calling "Gemini 3.5 Flash-Lite" by name, but the API reported the same model version, gemini-3.5-flash-lite, for both runs. So we read it for what it is: the same model asked the same question twice.

On one Arabic law question, the first run named no firm ("as an AI, I cannot favour one firm over another outright") and the second named five. Here are the other 19. For the Arabic questions, names are shown in the Latin form the answer gave in brackets.

The same business came first on 4 of 19 questions: 3 of 10 in English and 1 of 9 in Arabic. The two runs shared at least one name on 14 of 19, which was every English question (10 of 10) but only 4 of the 9 Arabic ones. Two of the "different" first picks look like the same brand under a slightly different name (Noưra, and Blossom), so a looser count is 6 of 19. Either way, the business at the top changed more often than it stayed.

Does a bigger model agree with a smaller one?

Not much more. We asked the first 10 English dental questions to Gemini 3.8 Flash, a bigger Gemini model, also from memory. It agreed with Flash-Lite on the first clinic once (Saudi German Hospital, for an emergency dentist near JLT) and shared at least one name on 4 of the 9 questions both answered.

The bigger model did handle one detail better. Asked about Al Nahda, which Arabic live search had confused with districts in Jeddah and Riyadh, it got the place right:

Why do the lists differ so much?

Our reading of the data, cause by cause:

• Different inputs. An answer from memory leans on a few well-known names; an answer from live search uses whatever pages fit the question. Only 20.3% of the names Gemini remembered appeared anywhere in Claude's live answers for the same industry (memory vs live search).
• Different pages. Each search finds a handful of pages, and 1,283 of the 1,954 sites Claude cited were cited only once (which websites AI cites). A different tool finds a different handful.
• Chance. The same model, asked the same question minutes later, changed its first pick on 15 of 19 questions. When many businesses fit a question, the assistant can pick a different set each time it writes the answer.
• Language. Arabic answers name fewer businesses and have fewer good pages to draw on, so there is less to agree on (Arabic answers).
• Spelling. The same business shows up as "Shuraa", "Shuraa Business Setup" and in Arabic script. Assistants, and the tools that track them, do not always connect these.

What does this mean for tracking AI visibility?

Your place in AI answers is a rate, not a rank. The useful number is: across my buyers' real questions, on several tools, in both languages, in what share of answers was I named, and who was named instead. Measured the same way each time, that number can move. A single "we're first on ChatGPT" screenshot cannot.

1. Fix a question list. 10 to 40 questions your buyers really ask, with the areas you serve, in English and Arabic. Keep it the same between checks.
2. Cover more than one assistant. Two tools shared a name on 16.4% of our questions, so each one needs its own check. At a minimum, ChatGPT and Gemini (Google AI).
3. Ask more than once. One run is noisy. Repeat the important questions and count how often you appear.
4. Count shares, not first place. Record the share of answers that name you and the competitors named instead.
5. Track Arabic separately. With 6.8% agreement in Arabic, an English result says nothing about the Arabic one.
6. Keep dated records. Save the raw answers or screenshots with the date and the tool, and re-check on a schedule.

This is how our plans work. Each one re-checks a fixed question list on ChatGPT and Gemini (Google AI): Foundation (AED 4,500 a month) 25 questions monthly in one language, Growth (AED 8,500) 40 English and 40 Arabic every 2 weeks, Authority (AED 16,000) 60 and 60 weekly. Plans have a 3-month minimum, then run month to month, with no setup fee. The free AI check is a quick first look.

What this data does not show

• One day, 21 October 2026. The main comparison used one run per question per lens.
• The main comparison pairs live search (Claude) with memory (Gemini), so it measures two different tools working two different ways, which is what buyers meet in practice but not a test of either alone.
• The repeat test (19 questions) and the 3.8 Flash test (10 questions) are small and mostly about Dubai Marina and dental clinics. They show direction only.
• Name matching misses some spelling variants, so all agreement figures are lower bounds.
• Neither lens is identical to what a person sees in the ChatGPT or Gemini apps.