Gemini isn't answering your question. It's routing you into Google.

Ask Gemini for the best boutique hotels in Dubai and you get Google’s own hotel inventory back. We asked it signed out, from the UAE, and it returned the three properties Google Search puts at the top of its Hotels module: Meliá Desert Palm, Andaz Dubai The Palm and The Canvas Dubai MGallery. Same ratings, same 5-star classification, and review counts matching to the unit. The order and the nightly rates differ, which tells you it is a separate query against the same inventory rather than a copy of the results page.

That pattern ran through the whole test. Gemini switched between four different Google properties depending on the category, and in the three categories where it didn’t, it cited nothing at all. The practical result for a Dubai business: before you publish anything for AI search, find out which Google feed owns your category, because on Gemini that feed is the answer.

This is the third of three pieces from our September test. Piece 1 looks at companies writing the articles AI quotes about them. Piece 2 measures how often the engines agree on who’s best.

What did we test?

We asked four surfaces the same question, “Best [category] in Dubai”, across seven categories: law firms, B2B SaaS, aesthetic clinics, boutique hotels, industrial manufacturers, universities and car dealerships. The surfaces were ChatGPT, Perplexity, Gemini and Google Search. Every run was logged out, on the free tier, from a UAE location. That gave 44 answers in total, 11 per surface, with three repeat runs on clinics and manufacturers. Full conditions and limits are in the methodology section at the end.

Which Google feed does Gemini use in each category?

Category	What Gemini showed	Google property behind it	Runs where it appeared
Boutique hotels	Hotels module with live nightly rates and review counts	Google Hotels	1 of 1
Aesthetic clinics	Place cards with ratings and opening hours	Google Maps / Business Profile	3 of 3
Industrial manufacturers	Product cards with merchant names and AED prices	Google Shopping	2 of 3
Universities	A YouTube video card	YouTube	1 of 1
Law firms	Text only, zero citations	None visible	1 of 1
B2B SaaS	Text only, zero citations	None visible	1 of 1
Car dealerships	Text only, zero citations	None visible	1 of 1

Four categories, four different Google properties. The other three got plain text with no sources at all.

Hotels: the answer is the Hotels module

Property	Google review count	Gemini review count	Rating
The Canvas Dubai MGallery	7.5K	7,540	4.4
Meliá Desert Palm	2.6K	2,588	4.6
Andaz Dubai The Palm	3K	3,033	4.6

Google rounds its review counts and Gemini gives the exact figure. Same three properties, same ratings, same 5-star classification, review counts matching to the unit.

The order and the nightly rates differ. Google returned Meliá, then Andaz, then Canvas, at AED 589, 320 and 212, with dates and a “Top-rated” filter applied. Gemini returned Canvas, then Meliá, then Andaz, at AED 188, 749 and 455, with no date context. So this is a separate query against the same inventory, not a mirror of the results-page module.

That distinction strengthens the finding rather than weakening it. Matching rates on a given date could be coincidence. Review counts matching to the unit across three properties cannot.

Google Search, signed out, UAE. The Hotels module sits at the top of the page.

Gemini, signed out, UAE. The order and the rates differ from Google’s module; the properties, ratings and review counts do not.

What moves a hotel up that answer is its Google Hotels listing and its rate, not an article about it.

Clinics: the answer is Google Business Profile

For aesthetic clinics, Gemini’s answer was built from place cards: name, rating, category, opening hours and address. No editorial sources and no citations. It was also the most stable answer in the study. Hortman Clinics was first and The Nova Clinic second in all three runs.

Manufacturers: product cards from Google Shopping

In two of three runs, Gemini added Google Shopping product cards to an answer about industrial manufacturers: three cards in the first run, eight in the third. The merchants were small local sellers with product feeds, including DANA (a water chiller at AED 11,000), ANBI Online, UaeGamer.com, shopURtool and bayanuae.com. None of them appeared in the text list above the cards, which named EGA, Ducab and Jubaili Bros.

Gemini, signed out, UAE. Merchant product cards inside a B2B manufacturer answer.

Two of three runs is repeatable but not guaranteed, and we only tested one product-adjacent category. Treat this as an open door worth checking, not a proven channel.

Universities: a YouTube card that also ranks on Google

Gemini surfaced a YouTube video card, “Top 10 Universities in Dubai for International Students” by R247 Success, 5.5K views. The same video ranks organically on Google’s results page for the same query. So does uniRank, the source ChatGPT used for its universities answer. Two engines, one results page underneath.

How much has Google’s AI Overview grown in Dubai?

It hasn’t grown. It’s already there, and that’s the more useful finding. An AI Overview fired on 4 of our 7 queries in September, and on the same 4 of 7 in our clean August baseline five weeks earlier. Same four categories both times.

Category	August, signed out	September
Law firms	Yes	Yes
B2B SaaS	Yes	Yes
Industrial manufacturers	Yes	Yes
Universities	Yes	Yes
Boutique hotels	No	No
Aesthetic clinics	No	No
Car dealerships	No	No

The four with an AI Overview are the categories people research before they buy. The three without are local or transactional, and there the local pack, the ads or the Hotels module fill the top of the page instead.

One number we are not publishing, and why. Our first August scan recorded an AI Overview on 1 of 7 queries, boutique hotels alone. That scan ran on a signed-in profile. We re-ran those cells signed out on 17 August and got 4 of 7 — and hotels, the only category the signed-in scan found, had lost its AI Overview. We use the signed-out run as the baseline because it is the one that matches September’s conditions. Comparing September against the signed-in figure would show coverage tripling in a month, which reads well and is an artefact of a sign-in change. Both runs are single samples, so we cannot separate personalisation from a platform change or from ordinary run-to-run variance.

Category	Named in the AI Overview	Sources it cited
Law firms	Al Tamimi & Company, Clyde & Co, Hadef & Partners	Legal 500 (×2)
B2B SaaS	Nazm.ae, Dukkantek, Verofax	F6S, DXB Start, Wellfound
Industrial manufacturers	Ducab, EGA, NAFFCO	Rotated between runs (see next section)
Universities	University of Birmingham Dubai, Canadian University Dubai	AECC Global, TopUniversities (QS), University of Birmingham’s own site

Google Search, signed out, UAE. The AI Overview on law firms is grounded on Legal 500.

For a Dubai business in one of those four categories, the practical reading is that the AI Overview is not a thing arriving later this year. It was sitting above your organic result in August and it was still there in September.

Can you get cited in an AI Overview on purpose?

You can get into the pool. You can’t pick the citation. We ran the manufacturer query three times. The AI Overview named the same three companies every time, and the organic results below it were identical every time. The sources it cited changed on every run.

Run	Brands named	Sources cited
1	Ducab, EGA, NAFFCO	mazeed (+1), Invest in Dubai, Scribd
2	Ducab, EGA, NAFFCO	Peko.one, Invest in Dubai, Scribd
3	Ducab, EGA, NAFFCO	Peko.one, Jafza, Peko.one

Google Search, signed out, UAE. AI Overview (top) citing mazeed, local pack below.

The organic page is stable. The AI Overview draws its citation from a small group of eligible pages and rotates between them. mazeed is also Perplexity’s main source for the same query, so one aggregator page was feeding two surfaces.

Within a single day the names held and only the sources moved. Over five weeks the names moved too. B2B SaaS had an AI Overview in both August and September, and between them the list it led with turned over almost completely: GrubTech, Dukkantek, Keepface and Saphyte in August; Nazm.ae, Dukkantek and Verofax in September. One company out of four survived the month. In manufacturers over the same period, EGA and Ducab held both times. Whether your category’s AI Overview names a stable set or reshuffles it is worth knowing before anyone sells you a position in it.

For a buyer, this is the useful part. Being one of the pages Google considers citable is winnable, and it’s mostly ordinary search work. Being the cited source on a given day isn’t something anyone controls. An agency promising you a specific AI Overview citation is promising something the data says doesn’t hold still.

What does ChatGPT read before it names a business?

Whatever authority exists in the category. Where there’s a directory, ChatGPT uses the directory. Where there’s trade press, it spreads across the trade press. Where there’s neither, it ranks on ratings and review volume and cites nothing against individual businesses.

Category	What grounded the answer	Businesses with their own citation	What moves the answer
Law firms	Legal 500, sole source	8 of 8	Directory standing
B2B SaaS	eChai Ventures, Analytics Insight, DXBStart	8 of 8	Earned trade coverage
Aesthetic clinics	Doctify, Practo, used for framing only	0 of 17 across three runs	Ratings and review volume

The format was the same throughout: text plus a comparison table. ChatGPT showed no map cards, no hotel module and no product cards in any category. It’s a text-and-citation surface, which makes its sources the thing to study.

Why do Perplexity and ChatGPT give different answers?

They read different kinds of source. Perplexity leaned on editorial coverage in almost every category: GRAZIA, Country & Town House, What’s On, Upgraded Points, Manufacturing Digital, Legal 500. ChatGPT leaned on listings and directories: Bayut, Best in Hood, Doctify, Dubai Industrial City, dubicars.com.

Perplexity also shows two answers at once on local queries, a Places module and a written answer, and they can disagree completely. For boutique hotels, the Places module led with Rove La Mer Beach (4.9, 14,933 reviews), DAMAC Maison Cour Jardin and Palace Downtown. The written answer recommended XVA Art Hotel, La Ville Hotel & Suites, Mazmi Casa, Beach Walk and Al Seef Heritage Hotel. None of the hotels in one appeared in the other. The Places module follows review volume; the written answer follows editors.

Piece 2 measures how far apart the engines land as a result.

What does this mean for a Dubai business?

Start with the feed, then the content. For most local businesses, the Google Business Profile is doing more AI work than the website.

Look at Dubai’s industrial local pack. The top three listings in September were Shuaiba Industrial Company, rated 4.5 on 35 reviews; NAZ Industries, 5.0 on 16; and Johar Manufacturing Services, 4.8 on 6. Johar, with six reviews, also appeared in ChatGPT’s manufacturer shortlist in two of three runs. Six reviews is the whole of third place in an entire category’s local pack. That is how little it currently takes to be visible here.

If your category looks like…	The engine answer is built from…	Where to start
Local service (clinics, trades, industrial)	Google Business Profile, ratings, reviews	Profile accuracy, categories and review volume
Hotels and stays	Google Hotels records and live rates	Hotels listing and rate parity
Products with a price	Google Shopping product cards (2 of 3 runs, one category tested)	A clean Merchant Center feed
Researched purchases (law, universities, B2B)	Directories, rankings and editorial coverage	Standing in the sources ChatGPT and AI Overviews cite

Gemini reads Google’s own properties, and ChatGPT and Perplexity keep citing pages that rank on Google. Search work still sits underneath all of it.

Methodology

• Prompt: “Best [category] in Dubai”, one prompt shape, English only.
• Categories: law firms, B2B SaaS, aesthetic clinics, boutique hotels, industrial manufacturers, universities, car dealerships.
• Surfaces: ChatGPT, Perplexity, Gemini, Google Search.
• Conditions: data collected 10 September 2026. Four surfaces, logged out, clean profile, geo verified to the United Arab Emirates, free or default model tier. Every cell was run in a fresh session. The logged-out Gemini tier served Flash-Lite.
• Volume: 44 answers. Aesthetic clinics and industrial manufacturers were run three times on every surface (n=3). The other five categories were run once per surface (n=1), so their answers could differ on a repeat run.
• Market: Dubai only. We don’t know whether these patterns hold in Abu Dhabi, Riyadh or anywhere else.
• August baseline: the AI Overview coverage table compares September against our 17 August re-run, which was signed out on the same conditions. An earlier 13 August scan of the same queries ran on a signed-in profile and is not used as the baseline; the difference between the two, and why it matters, is set out in the AI Overview section. Both runs are single samples, so personalisation, a platform change and run-to-run variance cannot be separated. The local pack figures are September only — the interval against any August baseline could not be established, so no change is claimed.
• What this doesn’t cover: signed-in or paid tiers, which can behave differently (signed-out Gemini returned no citations on non-local categories), other prompt phrasings, Arabic queries, and Claude and Copilot.
• Screenshots: recaptured after the logged run under the same conditions. AI answers can shift between runs, so every number in this article comes from the run log, and the screenshots illustrate the pattern.

---

Check your own category in two minutes. Open Gemini signed out and ask for the best [your category] in Dubai. If a Maps, Hotels, Shopping or YouTube block appears, that Google feed is where your visibility is decided, and it’s the first thing to fix. If you want a second opinion on what you find, email joe@brandure.io with the screenshot.