In a small study collected on 26 September 2026, ChatGPT and Gemini sometimes recommended the same Dubai supplier brands and sometimes returned entirely different shortlists. Both named Nook and Servcorp for a one-day desk near DMCC Metro. Their monthly car-rental answers shared no brands. For a seven-seat SUV collected at Dubai International Airport, both named Hertz, SIXT and Thrifty.
These are observations from six English questions and 12 completed answers, not a ranking of Dubai businesses. We tested existing signed-in consumer accounts, and two ChatGPT answers introduced a Lunasol reference that was absent from the questions. Account context was therefore a material limitation. Perplexity was initially planned but excluded after its first attempt required sign-in; no Perplexity recommendation results are reported.
Methodology at a glance
- Collection date: 26 September 2026
- Platforms: ChatGPT and Gemini
- Questions: 6
- Completed analysed answers: 12
- Language: English
- Runs per question/platform: 1 completed answer
- Conversation state: New conversation for every test
- Account state: Signed in
- Personalisation/memory: Not disabled or controlled
- Location: Not a controlled Dubai-IP test
- Arabic testing: Not included
- Perplexity: Excluded after access failure
- Purpose: Exploratory observation, not a representative ranking
Which businesses appeared?
Each cell below records the supplier brands offered as candidates in one completed answer. It includes qualified suggestions where the assistant attached a caveat. The names are observed recommendations, not Lunasol endorsements or independently verified confirmations that each supplier can fulfil the purchase.
| Buyer question | ChatGPT candidates | Gemini candidates |
|---|---|---|
| P1: One-day desk near DMCC Metro | Nook; Servcorp, Almas Tower; One Business Centre, One JLT | Nook; Servcorp, Almas Tower; The Place, Dome Tower |
| P2: Private office for six in Business Bay | Oh My Desk; Centrum Spaces, Bay Gate; Jetset Business Center, Prime Tower | UNBOX; Regus; Oh My Desk |
| P3: Monthly economy sedan delivered to Dubai Marina | SHIFT; Thrifty; Hertz | Quick Drive; Rentai; Rental Cars UAE |
| P4: Seven-seat SUV collected at DXB | Thrifty; Hertz; SIXT | Hertz; SIXT; Thrifty |
| P5: Try ergonomic chairs in a Dubai showroom | BAFCO; OfficeMaster; WORKSPACE | Ergospace; Haworth; ABC, supplying Herman Miller |
| P6: Ten workstations with delivery and installation | Highmoon; Mahmayi; MR Furniture | Highmoon; Office Plus; Office Inn |
The airport row agrees at brand level, not at branch level. ChatGPT displayed Terminal 3 entries for Thrifty and Hertz and a Terminal 1 SIXT entry, alongside links discussing other terminals. Gemini displayed Terminal 1 entries. A traveller still needs to check the actual collection counter, vehicle and dates.
Across the 12 answers, there were 36 candidate appearances and 27 distinct normalised supplier brand names. Those are not 27 verified independent legal businesses. Corporate affiliations were not audited, and different brands can be related. No business was awarded a visibility score by pooling unrelated sectors.
What differed between the two sets of answers?
Each row below compares the two captured answers to the same question. The final column counts distinct candidate brand names across that pair, counting a shared brand once. It is not a count of independently verified legal businesses or a visibility score.
| Prompt | Shared brands | Total distinct candidates across both answers |
|---|---|---|
| P1: One-day desk near DMCC Metro | Nook; Servcorp | 4 |
| P2: Private office for six in Business Bay | Oh My Desk | 5 |
| P3: Monthly economy sedan delivered to Dubai Marina | None | 6 |
| P4: Seven-seat SUV collected at DXB | Hertz; SIXT; Thrifty | 3 |
| P5: Try ergonomic chairs in a Dubai showroom | None | 6 |
| P6: Ten workstations with delivery and installation | Highmoon | 5 |
These observations do not estimate how frequently a brand would appear in repeated searches. The row totals should not be added to obtain a study-wide unique-brand count, because brands can recur in different questions.
The most useful distinction is between a broad category and a purchase constraint. A desk near DMCC Metro, a six-person office in Business Bay and ten installed workstations are different buying needs. Their answers should be assessed against those requirements, rather than combined into a citywide league table.
Agreement can hide different purchasing details
The airport answers shared all three brands, but the branches, order and wording differed. Agreement on names did not establish agreement on collection arrangements or live inventory. Neither answer constitutes a booking confirmation.
A shortlist can include unresolved fit
ChatGPT offered One Business Centre for the one-day desk question while flagging uncertain immediate day-pass availability. It also distinguished Servcorp's membership structure from a straightforward one-off pass. These remained candidates under our counting rule because the answer actively proposed them.
For the workstation question, ChatGPT included MR Furniture while saying it had not found an exact published ten-workstation package. Gemini included Office Inn while noting that it had not independently verified a physical showroom through Maps. The latter caveat did not exclude the supplier because P6 asked for delivery and installation, not a showroom visit.
These examples explain what the table counts. They do not resolve the assistants' underlying claims about supplier suitability.
A new conversation did not remove account context
In P5, ChatGPT referred to chairs for the Lunasol office or team. In P6, it referred to a Lunasol-type office. Neither prompt mentioned Lunasol. We retained both first completed answers rather than replacing them with cleaner-looking results.
The mechanism behind those references was not established. Account context or personalisation was a plausible confounder that we did not control. We did not inspect or disable account memory or other personalisation settings. Consequently, differences between the two products cannot be attributed solely to their models. The absence of an obvious personalised phrase in another answer would not prove that answer was unaffected by context.
What did checking the sources reveal?
We checked the named suppliers' own websites to establish their public brand identities and Dubai relevance. That is a narrower check than verifying every recommendation, price, warranty, opening time or available vehicle. Supplier websites are also first-party statements, not independent quality assessments.
One concrete discrepancy appeared in the coworking results. ChatGPT described Nook as open every day, including weekends. When inspected on 26 September, Nook's own website listed Sunday as closed. This records a conflict between the answer and the website, not a definitive finding about real-world opening hours. A customer planning a Sunday visit would need to resolve it directly.
The captured answers also used different source types. The coworking answers included links to booking platforms such as Letswork, Deskimo and Hotdesk, alongside supplier pages. Those publishers were not counted as recommended coworking operators merely because they supplied a link.
We retained visible source links, but some grouped additional citations were not expanded. This study therefore does not offer a complete citation inventory or percentages showing which source types each platform prefers. A visible link is evidence of what the interface displayed, not a full record of the system's retrieval or selection process.
Exact questions and collection method
The six questions were selected before collection to cover two practical buying needs in each of three sectors: coworking, car rental and office furniture. This was a handpicked exploratory sample, not a random sample of Dubai demand.
The full question for each run consisted of the relevant text below followed by the same instruction shown afterwards:
- P1: Which coworking spaces would you recommend for a one-day desk near DMCC Metro in Dubai?
- P2: Which coworking providers would you recommend for a private office for six people in Business Bay, Dubai?
- P3: Which car rental companies would you recommend for a monthly economy sedan with delivery to Dubai Marina?
- P4: Which car rental companies would you recommend for a seven-seat SUV rental with collection at Dubai International Airport?
- P5: Which office furniture businesses would you recommend for trying ergonomic desk chairs in a Dubai showroom before buying?
- P6: Which office furniture suppliers would you recommend for ten workstations with delivery and installation in Dubai?
Shared instruction: “Search the web and recommend up to three businesses. Briefly explain the fit and link to sources. If you cannot verify a fit, say so.”
Collection took place on 26 September 2026, approximately 14:47–15:06 UTC, through consumer web interfaces. ChatGPT was signed in on a Pro account with the visible “Extra High” setting; the underlying model identifier was not established. Gemini was signed in with the visible “Flash” setting. No API responses were substituted.
Each question started a fresh conversation. Existing account settings were retained. Gemini's interface showed Warsaw, Poland, as its IP-based location, so this was a study of Dubai questions, not a controlled Dubai-IP test. It did not cover Arabic questions or establish how a Dubai resident's account would respond.
We retained the first completed answer for each question and product. Gemini's initial P3 attempt returned a technical error; the predeclared procedure allowed one retry in a fresh conversation, which completed. The error was retained separately and did not become an extra answer. Perplexity's sign-in block was an access failure, not a valid answer recommending zero businesses. Collection continued with ChatGPT and Gemini after that block.
The final analysed dataset contains six ChatGPT answers and six Gemini answers. Answer text and visible links were retained internally, with private account navigation excluded from the research records. The table and prompt list provide the public observation summary; repeating them today may produce different results.
How names were counted
A recommendation meant a supplier actively offered as a candidate, including one with qualifications. Incidental comparisons, rejected alternatives and source-only publishers were excluded. A repeated name counted once per answer.
Clear spelling variants were normalised, while relevant branch labels were retained. ABC supplying Herman Miller counted as one supplier entry, not two. Rental Cars UAE and the answer's associated Amex name were treated as one entry. Product names did not become extra suppliers. This is brand-level analysis rather than an audit of corporate ownership.
What should a Dubai business take from this?
Use a captured recommendation as a starting point for investigation. Check whether the answer describes the correct offer, branch, service area and buying conditions. A mention can be commercially unhelpful if it sends someone to the wrong location or promises an unavailable arrangement.
This sample cannot tell a business how often it appears across AI search, which optimisation caused an appearance, whether a missing brand is generally invisible or which assistant produces the best answers. It contains no repeated-run stability estimate, customer traffic measurement or lead attribution. The three selected sectors do not represent Dubai as a whole.
A useful next step is to test your own important buying questions over time and inspect the evidence behind both appearances and omissions. Keep those observations separate from any claim that a particular website edit caused a change.
Lunasol publishes this study and offers commercial GEO services. The observations are not evidence that Lunasol caused any supplier's appearance. If you need to investigate your own questions, competitors and public information, Lunasol's GEO service provides a relevant starting point for defining that work.