Do ChatGPT, Gemini and Google AI Mode name the same Dubai brokerages?

01The result

What wave 2 returned, against a threshold set before its numbers existed

The verdict on the pre-registered basis

A second basis, with Google AI Mode's answer cards, carries no verdict

The headline figure is 90.4%, on the pre-registered basis above, across 40 Dubai property questions. On a second basis that also reads Google AI Mode's answer cards, added after registration and described below, 110 of 117 pairs were below the mean (94.0%), median ratio 0.662, over the same 40 questions, with 3 pairs excluded: G4 on all three engine pairs. That figure carries no verdict.

The threshold was set for the pre-registered basis and is not applied to any other. Measured against the same threshold, the extended figure falls on the same side of it. The channel decision described below changes the magnitude slightly and does not change the result.

Google AI Mode reported no model identifier

Two of the three engine pairs include Google AI Mode, which reported no model identifier on any of its 800 answers. Whatever it returned could reflect a model change that cannot be seen from outside, and every figure with AI Mode answers inside it carries that.

All three verdict paragraphs were in the remote before the statistic's first recorded run on these rules

Three verdict paragraphs, one for each outcome the threshold allows, were committed together in 889d568 at 04:05:40 UTC on 4 October and pushed at 04:37:58. The statistic's first recorded run on the rules this page uses came 33 hours later, at 13:48:06 UTC on 5 October, and its verdict line selected the one above.

All three were in the remote before that run. They were not written before every wave 2 figure: a session printed headline figures on 25 September, with no rulings applied, as the pre-registration section below discloses.

The other two were deleted in fdce211, a commit that only deletes lines, committed at 08:54:16 UTC on 6 October and pushed at 09:08:59.

An independent reimplementation found a defect; both figures and the verdict are unchanged

An independent reimplementation, written from the committed procedure by a session barred from the original code, found a defect in it. gate.py resolved the first form it saw of each name and applied that result to every other form of the same name, so 13 mentions in 13 wave 2 runs, and 4 in wave 1, took another form's result.

The defect was fixed and the statistic recomputed on 6 October. Both figures and the verdict are unchanged, and one question's row in the per-pair table, F4, moved. The verdict paragraph published here was selected from the first run; the corrected run selects the same one.

With the engine label shuffled, 27.6% of pairs fell below the mean on average

The statistic was also run with the engine label shuffled across runs, everything else held fixed, so that the label carries no information. Across 1,000 shuffles under seed 20260923, the labels shuffled within each question, the share of pairs below the mean averaged 27.6%, with the middle 95% of shuffles between 18.8% and 37.6%, and the median ratio averaged 1.103, between 1.055 and 1.150.

None of the 1,000 reached the real 90.4% or 0.652.

The shuffled median ratio sits above 1 because each half of an engine's agreement with itself is the union of 10 runs, while the overlap between engines compares unions of 20, and larger unions overlap more. The shuffle keeps that difference, so it is the baseline that controls for it.

That says what the machinery returns when the engine makes no difference. It does not say whether a difference between engines is in which firms they name or in how each one spells them, so the spelling check is reported separately.

Spelling could account for at most 1.7 percentage points of the share, by this test

Of the 188 firms resolved to one licence and named at all, 66 were named on more than one engine. Each of those was spelled 2.35 ways on average within the engine that spelled it most variously, and 3.00 ways across all three.

Crediting every spelling a generous match accepts as the same firm gives 102 of 115 pairs below the mean (88.7%), median ratio 0.662, against 104 of 115 (90.4%) and 0.652 strictly, so in wave 2 spelling could account, by this test, for at most 1.7 percentage points of the share and 0.010 of the median ratio; abbreviations and trade names it cannot see are outside it.

Wave 1 as a repeat measure

Wave 1 asked the same 40 questions, word for word, on 19 September. It is published alongside as a repeat measure and never combined with wave 2 into any figure.

Drift check: ChatGPT and Gemini agree with wave 1, and Google AI Mode diverges

Its comparison with wave 2, cell by cell, was fixed before wave 2 ran and is reported whatever it shows. A cell diverges when its overlap across the waves is more than 0.10 below the lower of its split-halves, and an engine is called diverged when at least a quarter of its cells do.

On the 14 questions wave 1 asked on all three engines (G4 aside, which names no firm), that is 4 of 14: ChatGPT had 0 diverging cells, Gemini 2 (D3, D6) and Google AI Mode 4 (A1, C1p, C4p, D5). ChatGPT and Gemini agree with wave 1, and AI Mode diverges. AI Mode reports no model identifier, so its divergence may be a model change nobody outside can see.

On the 25 questions wave 1 asked on ChatGPT alone the bar is 7, and 1 cell diverged (A2), so ChatGPT agrees.

Drift and replication are two different claims

These are two different claims. The replication is about the pattern across engines: in both waves, two engines answering the same question named fewer firms in common than each did with itself.

The drift check is about which firms one engine names from 19 to 23 September. AI Mode naming different firms in wave 2 does not contradict the pattern repeating; its list moved while the pattern held.

On wave 1's 15 three-engine questions, wave 2 gives 37 of 42 pairs below the mean

On the 15 questions wave 1 asked on all three engines, wave 2 gives 37 of 42 pairs below the mean (88.1%), median ratio 0.642, with G4's 3 pairs excluded. That is the same count as wave 1's, with a median ratio of 0.642 against wave 1's 0.656. It was computed after the headline, at the pre-publication review's suggestion, is not pre-registered, and is reported as found.

Both waves were re-extracted and re-resolved under the same rules for this comparison, which moves wave 1's own headline: before re-extraction it was 37 of 42 pairs below the mean (88.1%), median ratio 0.685, and after it, under the rules as corrected on 13 October, 37 of 42 (88.1%), median ratio 0.656. The share did not move and the ratio moved by 0.029 in two steps: re-extraction took it from 0.685 to 0.676, and removing Arjan took it from 0.676 to 0.656 (Corrections).

02Pre-registration

What was fixed before the numbers existed

Times are UTC. "Pushed" times are the remote's own record of when the commit reached GitHub. Commit times are written by our machine and are only as good as its clock.

The threshold landed 30 minutes into the collection, not before it. No answer had been extracted and no result existed when it was committed, so it was fixed before the numbers existed. It was not fixed before the run started, and we do not describe it that way.

Wave 1 was collected on 19 September and asked 25 of its 40 questions on ChatGPT alone, so every comparison between engines rested on 15 questions. Wave 2 asked all 40 on all three engines, for coverage and for no other change. The recollection was decided and pushed before wave 2 ran.

Wave 1's result was known when wave 2 was designed, and that is why the threshold for calling it replicated was committed before wave 2's numbers existed: what was registered is a procedure, not a blind hypothesis.

A threshold written in advance protects the headline number. It does not protect the judgement calls downstream of it, which anomaly gets chased or which matching edge case resolves which way, and three further guards were written for those. They are described under the method below.

The procedure committed that wave 1's payloads and ledger would stay as they are. The payloads did: all 1,400 match the SHA-256 recorded at collection.

The ledger did not: it was re-extracted so that both waves are resolved by the same rules, as the same procedure requires, and redacted, because the person rulings required it. Wave 1's figures are published from before and after: 37 of 42 below the mean both times, median ratio 0.685 before and 0.656 after re-extraction and the 13 October corrections.

The procedure committed that the permutation null would run before the real statistic. Its specification was committed at 77214fb, 03:11 UTC on 23 September, before any wave 2 figure existed, and its first recorded run was on 6 October, after the headline ran on 5 October. The protection the commitment was written for is in the specification rather than the run order: the null's form could not be chosen to fit a result nobody had seen.

One change to the basis after registration

Google AI Mode's answer cards entered the basis after registration, and the headline is computed without them

The basis committed at 17e8485 is "register-resolved firms, text channels, every run of the cell". In the code at that commit, the channels that name firms were two: prose_list, the leads of list items, and table_cell, the cells of tables (NAMING_CHANNELS in packages/geocensus/geocensus/extract/mentions.py).

Cards occur on AI Mode only, so the two bases differ on the two engine pairs that include AI Mode.

A rule on the pre-registered channels, written after a figure had been seen

On the code before R6, R7 and every later amendment and ruling, wave 2 gives 105 of 115 rather than 104, and the verdict is the same

A fourth check, added on 13 October, recomputes the headline under a rule read the way we did not read it. Read literally, the rule that a hand class gives way to the register would credit Sobha to its dormant entry in 23 wave 2 runs across 5 cells and 12 wave 1 runs across 2, and bare "Vacationer" in 1 wave 2 run, and would credit DAMAC to nothing. Wave 2 then gives 104 of 115 pairs below the mean (90.4%), median ratio 0.649 rather than the published 0.652, with the same verdict, and wave 1 gives 37 of 42 (88.1%), median ratio 0.649 rather than 0.656. The reading we took is the one that leaves the published figures as they are.

A session that saw figures

Headline figures were printed on 25 September, before R6 and R7 were written, and no figure here is computed from that print

Between 20:58 and 21:07 UTC on 25 September (04:58 to 05:07 on 26 September in our own timezone, UTC+8), the session writing the extraction rules ran a wave 2 script that printed headline figures on resolution with no rulings applied. The script writes only to the screen: no commit records those figures, and the disclosure does not give them.

The first wave 2 headline figure in the repository was committed at 13:48:17 UTC on 5 October, by a later run, and no figure on this page is computed from that print. The figure before R6 and R7 is computed, by a later run, from the table gate.py wrote on 23 September alongside the named sets that script read.

The same session wrote R6 and R7 about eleven hours later. The section above, "A rule on the pre-registered channels, written after a figure had been seen", gives that sequence, and we do not claim the figures had no influence on the rules it wrote.

The run was disclosed in the committed rules at the time, and it is the reason the implementation went through the review rounds described below.

A run that wrote the statistic's inputs

The statistic's inputs existed from 23 September, and no record shows the statistic computed that day

03Method

How it was measured

What was asked, of which engines, and when

Between 02:36:12 and 05:41:21 UTC on 23 September 2026 we put 40 questions about buying, selling and letting property in Dubai to ChatGPT, Gemini and Google AI Mode, 20 times each, from a United Arab Emirates location: 2,400 answers, every question on every engine.

We extracted the firms each answer named, checked every one against the Dubai Land Department register, and asked one question of the result: when two engines answer the same question, do they name the same firms more or less than one engine agrees with itself?

The answer to that question was judged against a threshold committed to git while the collection was running and before any name had been extracted. It is above, first, whichever way it came out.

The questions

40 questions, listed word for word at the foot of this page: a general question, five budget bands, eight communities, eight property types, questions about selling, letting and keyword phrasing, and one probe asking whether bigger brokerages get named more often by AI. The pack is dubai-02-v2, hash a985350a5cda3b7a7ca86bc53aa958c192d1ab988dde1816034d65bf13098a04. Its 40 question texts are the ones wave 1 asked.

The engines

Three engines, reached through DataForSEO's API: ChatGPT with web search forced, Gemini, and Google AI Mode. Every request carried the same parameters: location United Arab Emirates (DataForSEO location 2784), language English, desktop, Windows.

ChatGPT reported a model identifier on 19 of its 800 answers (gpt-5-6) and none on the rest. Gemini reported 3.5 Flash-Lite on all 800. Google AI Mode reported none on any. These are the answers each product gave through that API on that day, not a claim about any underlying model.

The runs

40 questions on 3 engines is 120 cells, and each cell was asked 20 times: 2,400 runs in one window of just over three hours inside one UTC day. Every one of the 120 cells holds 20 runs. No request came back empty and every run has answer text.

The collection cost $9.60 for 2,400 paid requests, against a cap of $12.00 set before it began.

Repeated answers: collapsed to one run per distinct answer, wave 2 gives 100 of 111 pairs below the mean

Repeated answers are counted like the rest. 67 of wave 2's 2,400 runs repeat another run's answer word for word, every one inside its own cell: four Google AI Mode cells (B1, C1p, C2p, D7) returned one answer 17 to 19 times in 20.

A repeated answer raises a cell's agreement with itself, which pushes its pairs below the mean. Collapsed to one run per distinct answer, wave 2 gives 100 of 111 pairs below the mean (90.1%), median ratio 0.656; 4 more pairs are excluded because B1 and C1p are left with fewer than 4 runs and so have no split-half.

All four were below their split-half mean as run, so excluding them lowers the share rather than raising it; counted as not below, wave 2 would be 100 of 115 (87.0%), still above the 70% the threshold requires. Wave 1's 52 repeats, in three AI Mode cells, give 33 of 38 (86.8%), median ratio 0.661, against 37 of 42 and 0.656 as run.

R2 moves credit in both directions, and with it switched off the headline is one pair stronger

R2 re-points credit from the firm a text names to another firm when a listing card's contact number says so, and it moves credit in both directions: in wave 2, 2 firms gain 232 run credits and 6 lose 260, and 188 firms are named anywhere with R2 against 192 without.

With R2 switched off the headline is one pair stronger, 105 of 115 pairs below their split-half mean (91.3%) against the published 104 of 115 (90.4%), with the same median ratio of 0.652 and the same 5 exclusions. The committed threshold selects the same verdict either way. The published figure is the one computed under the pre-registered rules, with R2 in force.

The register

Every named string was checked against the Dubai Land Department's broker register, from the snapshot of 16 September 2026 (export SHA-256 20c008c3d48942a3…, firm table SHA-256 1b4546fbbd28340a…; full hashes in the bundle). A firm counts as named only when it resolves to a firm on that register.

A listing card whose contact number points to a different registered firm is re-mapped to that firm or dropped, under R2, committed on 19 September. A card with no signal either way keeps its name match: of the 342 registered firms matched from a listing card in wave 2, 220 were name-only, 112 were supported by a contact number, 3 were re-mapped and 7 were left unmatched. The rule also admits website and address evidence checked by hand; none was recorded in this edition.

Resolving a named firm to the register by its name is the ordinary path under the committed rules, and a listing card's contact number is an additional signal rather than a requirement: of the 188 firms counted in wave 2, 67 rest on a listing card's contact number, 65 where it supports the name match and 2 where it pointed from another firm's name to this one, and 121 rest on the name alone: 68 were matched from a card that gave no signal either way, and 53 were never matched from a card.

The register and the answers are the only sources. No property portal was fetched.

Two committed rules, R5 and trade-name matching, are not implemented in this edition

Of the eight extraction and resolution rules committed before wave 2 was collected, two are not implemented in this edition: R5, number-only resolution, and trade-name matching. A third, R2, ran on contact numbers only; its hand-checked signals were not recorded.

R5, committed on 21 September before wave 2 was collected, lets a listing card's contact number alone resolve a firm that no name reached, in a class of its own. It applies to wave 2 and is not implemented in this edition.

No row is excluded by it and no figure on this page depends on it: the pre-registered naming channels are prose and table cells, and a listing card enters the figures only through R2's check on a card already matched by name, which a number-only row by R5's terms is not. It is carried to the next edition.

The statistic

For each question and each pair of engines, the overlap of the two sets of firms named across all 20 runs, as a Jaccard index, is compared with how well each engine agrees with itself: the overlap of its first 10 runs with its last 10. A pair is below its own noise when the overlap between engines is lower than the mean of the two engines' self-agreement.

The two reported figures are the share of pairs below that mean and the median ratio of the overlap between engines to it. A pair where that mean is zero or undefined is excluded and counted, never dropped silently.

The script is analysis/dubai-02-wave2/headline.py, committed with the threshold.

Three guards on the judgement calls

The name-matching review of wave 2 was blinded: each row a pair of strings and a candidate licence number, no engine label, nothing showing which way a merge would move any figure, in an order shuffled under a committed seed. Wave 1's review was not blinded.

The permutation null above was committed before it was run on the real labels.

The headline was recomputed from the stored answers by a session briefed from the procedure alone, without wave 1's result, its direction or its size, and barred from gate.py, headline.py, every analysis output and the repository's history; it reported opening none of them. It returned 104 of 115 pairs below the mean (90.4%), median ratio 0.65, with the same five pairs excluded, against headline.py's 104 of 115 and 0.652. No pair is below the mean in one and not in the other, so both figures and the verdict are the same.

Underneath, they differ in one cell: T3 on ChatGPT, where the reimplementation credits one more firm. Its name is made only of category words and appears in prose and also as a map listing; the committed rule resolves each channel on its own, so the listing does not count for the prose mention. Three split-halves also differ by 0.01, from rounding a tie up rather than to the even digit.

The recomputation found a defect in gate.py, described with the result above.

Some strings are classed by hand rather than by a rule: a place, a developer, a portal or a holiday-homes operator is ruled not to be a brokerage whatever the register says. The rulings ship in registry/name_classes.csv in the bundle. Since 13 October 2026 a string can also be ruled unresolved: it may name a firm, but the register cannot say which, so it takes no licence and claims none in its place. That class was added for one string, a common word that begins 46 registered names, and it removes a credit by hand, so every string ruled this way is listed here and in Corrections.

Strings ruled unresolved: "URBAN".

A hand class gives way to the register only where the register holds an active licence under the brand's name; a dormant entry does not displace it. This was settled on 13 October 2026, in October about data collected in September, for the three hand-classed brands the register holds only as dormant entries: Sobha (SOBHA L.L.C, ORN 18609), DAMAC (DAMAC PROPERTIES CO (L.L.C), ORN 1590) and Vacationer (VACATIONER REAL ESTATE L.L.C, ORN 31359). Each keeps its class. We read an engine writing "Sobha" in answer to a question about agencies as naming the developer, and a dormant entry has no operating brokerage behind it. The figures under the other reading are given with the sensitivity checks.

04Firms by question

Which firms each engine named

Rows are ordered for reading, and the order is not a ranking. The table of the 20 most-named firms, below the findings at the top of the page, is ordered by the runs naming each firm; that order is a count, not a ranking of quality.

Each table shows at most 12 firms: the registered firms any of the three engines named for that question, ordered by the runs naming them summed across the engines, ties by name. A cell reads k/n: named in k of the cell's n runs.

Each firm is shown by two names: first the form the engines most often wrote for it, counted over every run, on any of the three engines and any of the 40 questions, in which a mention credited it; then the name the Dubai Land Department register holds for the firm those mentions were credited to. The engine form may carry a city or branch qualifier, such as Dubai or a branch name, where that is how the engines wrote it.

A1 Which are the best real estate agencies in Dubai?

29 firms in the union, 12 shown, ordered by runs named across engines.

A2 I want to buy property in Dubai. Which real estate brokerages should I work with?

27 firms in the union, 12 shown, ordered by runs named across engines.

Credited by R2: the engines wrote Christie's International Real Estate, and the contact number on the listing cards beside those answers is registered to Premier Estates' brokers. See "What the figures count".

A3 Which real estate brokerage in Dubai should I use to sell my apartment?

26 firms in the union, 12 shown, ordered by runs named across engines.

A4 Who are the most trusted real estate brokers in Dubai?

26 firms in the union, 12 shown, ordered by runs named across engines.

A5 I'm moving to Dubai and need to rent. Which agency should I go to?

20 firms in the union, 12 shown, ordered by runs named across engines.

B1 I want to buy an off-plan apartment in Dubai. Which real estate brokerages should I work with?

27 firms in the union, 12 shown, ordered by runs named across engines.

B2 I want to buy a ready apartment on the secondary market in Dubai. Which real estate brokerages should I work with?

31 firms in the union, 12 shown, ordered by runs named across engines.

B3 I want to buy a villa in Dubai. Which real estate brokerages should I work with?

33 firms in the union, 12 shown, ordered by runs named across engines.

Credited by R2: the engines wrote Christie's International Real Estate, and the contact number on the listing cards beside those answers is registered to Premier Estates' brokers. See "What the figures count".

B4 I want to buy a townhouse in Dubai. Which real estate brokerages should I work with?

28 firms in the union, 12 shown, ordered by runs named across engines.

B5 I want to buy a branded residence in Dubai. Which real estate brokerages should I work with?

31 firms in the union, 12 shown, ordered by runs named across engines.

Credited by R2: the engines wrote Christie's International Real Estate, and the contact number on the listing cards beside those answers is registered to Premier Estates' brokers. See "What the figures count".

B6 I want to buy an office or retail unit in Dubai. Which real estate brokerages should I work with?

37 firms in the union, 12 shown, ordered by runs named across engines.

B7 I want to buy a plot of land to build on in Dubai. Which real estate brokerages should I work with?

48 firms in the union, 12 shown, ordered by runs named across engines.

B8 I want to buy a short-term rental / holiday home investment in Dubai. Which real estate brokerages should I work with?

37 firms in the union, 12 shown, ordered by runs named across engines.

C1p I want to buy an apartment under AED 1 million in Dubai. Which real estate brokerages should I work with?

26 firms in the union, 12 shown, ordered by runs named across engines.

C2p I want to buy a property around AED 2 million (Golden Visa threshold) in Dubai. Which real estate brokerages should I work with?

34 firms in the union, 12 shown, ordered by runs named across engines.

C3p I want to buy a home in the AED 5 to 10 million range in Dubai. Which real estate brokerages should I work with?

29 firms in the union, 12 shown, ordered by runs named across engines.

Credited by R2: the engines wrote Christie's International Real Estate, and the contact number on the listing cards beside those answers is registered to Premier Estates' brokers. See "What the figures count".

C4p I want to buy a luxury property above AED 20 million in Dubai. Which real estate brokerages should I work with?

23 firms in the union, 12 shown, ordered by runs named across engines.

Credited by R2: the engines wrote Christie's International Real Estate, and the contact number on the listing cards beside those answers is registered to Premier Estates' brokers. See "What the figures count".

C5p I want to buy the cheapest property that still gets me residency in Dubai. Which real estate brokerages should I work with?

40 firms in the union, 12 shown, ordered by runs named across engines.

D1 I want to buy in Dubai Marina in Dubai. Which real estate brokerages should I work with?

17 firms in the union, 12 shown, ordered by runs named across engines.

D1-nat I want to buy property in Dubai Marina. Which real estate brokerages should I work with?

16 firms in the union, 12 shown, ordered by runs named across engines.

D2 I want to buy in Downtown Dubai in Dubai. Which real estate brokerages should I work with?

30 firms in the union, 12 shown, ordered by runs named across engines.

Credited by R2: the engines wrote Christie's International Real Estate, and the contact number on the listing cards beside those answers is registered to Premier Estates' brokers. See "What the figures count".

D3 I want to buy on Palm Jumeirah in Dubai. Which real estate brokerages should I work with?

23 firms in the union, 12 shown, ordered by runs named across engines.

D4 I want to buy in Jumeirah Village Circle in Dubai. Which real estate brokerages should I work with?

33 firms in the union, 12 shown, ordered by runs named across engines.

D5 I want to buy in Business Bay in Dubai. Which real estate brokerages should I work with?

25 firms in the union, 12 shown, ordered by runs named across engines.

Credited by R2: the engines wrote Christie's International Real Estate, and the contact number on the listing cards beside those answers is registered to Premier Estates' brokers. See "What the figures count".

D6 I want to buy in Dubai Hills Estate in Dubai. Which real estate brokerages should I work with?

24 firms in the union, 12 shown, ordered by runs named across engines.

D7 I want to buy a villa in Arabian Ranches in Dubai. Which real estate brokerages should I work with?

21 firms in the union, 12 shown, ordered by runs named across engines.

D8 I want to buy in Emirates Hills in Dubai. Which real estate brokerages should I work with?

26 firms in the union, 12 shown, ordered by runs named across engines.

Credited by R2: the engines wrote Christie's International Real Estate, and the contact number on the listing cards beside those answers is registered to Premier Estates' brokers. See "What the figures count".

E-a I want to buy an investment property in Dubai. Which real estate brokerages should I work with?

38 firms in the union, 12 shown, ordered by runs named across engines.

Credited by R2: the engines wrote Christie's International Real Estate, and the contact number on the listing cards beside those answers is registered to Premier Estates' brokers. See "What the figures count".

F1 I want to sell my villa quickly in Dubai. Which real estate brokerages should I work with?

28 firms in the union, 12 shown, ordered by runs named across engines.

Credited by R2: the engines wrote Christie's International Real Estate, and the contact number on the listing cards beside those answers is registered to Premier Estates' brokers. See "What the figures count".

F2 I want to sell an off-plan unit before handover in Dubai. Which real estate brokerages should I work with?

40 firms in the union, 12 shown, ordered by runs named across engines.

F3 I want to let out my apartment to long-term tenants in Dubai. Which real estate brokerages should I work with?

26 firms in the union, 12 shown, ordered by runs named across engines.

F4 I want to put my apartment on short-term rental in Dubai. Which real estate brokerages should I work with?

15 firms in the union, 12 shown, ordered by runs named across engines.

F5 I want to find a family home to rent with schools nearby in Dubai. Which real estate brokerages should I work with?

27 firms in the union, 12 shown, ordered by runs named across engines.

F6 I want to exit a property I bought off-plan that has not appreciated in Dubai. Which real estate brokerages should I work with?

39 firms in the union, 12 shown, ordered by runs named across engines.

F7 I want to buy a property with my pension as a retiree in Dubai. Which real estate brokerages should I work with?

35 firms in the union, 12 shown, ordered by runs named across engines.

Credited by R2: the engines wrote Christie's International Real Estate, and the contact number on the listing cards beside those answers is registered to Premier Estates' brokers. See "What the figures count".

F8 I want to buy investment properties as a portfolio, not one unit in Dubai. Which real estate brokerages should I work with?

42 firms in the union, 12 shown, ordered by runs named across engines.

G4 Do bigger Dubai brokerages get named more often by AI than small ones?

0 firms in the union, 0 shown, ordered by runs named across engines.

Asked 20 times on each of ChatGPT, Gemini and Google AI Mode on 23 September 2026, "Do bigger Dubai brokerages get named more often by AI than small ones?" (G4) drew no registered firm from any engine.

T1 best dubai real estate agency

26 firms in the union, 12 shown, ordered by runs named across engines.

Credited by R2: the engines wrote Christie's International Real Estate, and the contact number on the listing cards beside those answers is registered to Premier Estates' brokers. See "What the figures count".

T2 dubai villa broker

30 firms in the union, 12 shown, ordered by runs named across engines.

T3 jvc property agent

18 firms in the union, 12 shown, ordered by runs named across engines.

Names we could not match

Wave 2's answers held 3,480 distinct names that resolved to no registered firm, 175 of them named often enough that they would have been published had they resolved: in at least 3 of one cell's 20 runs, or in at least 5 runs across the edition, a floor set on 19 September 2026 before wave 2 was collected.

Over all 40 questions on each engine: 1,116 on ChatGPT (46 of them that often), 1,757 on Gemini (122) and 816 on Google AI Mode (100). A name can go unresolved on more than one engine, so the engines' counts add to more than the total. Wave 1's answers held 2,046 (43).

A string that did not resolve was not matched by any method available to this study. That is a statement about our matching, not about whether the firm holds a licence.

Individuals

These answers name people as well as firms. Wave 2's answers name at least 154 distinct individuals: 52 on ChatGPT, 79 on Gemini and 32 on Google AI Mode.

Counted per engine that is 163, more than 154, because some individuals are named on more than one engine. Wave 1's name at least 119.

Individuals are counted and never named on this page or in the bundle, and no count implies anything about whether a person is licensed. Every person count is a floor, for the reasons under Limitations.

Our own site

We publish research about AI visibility, so we checked whether the engines cite or name us. On G4, the question about AI visibility, ChatGPT cited realestateseo.ph in 5 of 20 runs in each wave, Google AI Mode in 4 of 20 in wave 2 and none in wave 1, and Gemini in none. No engine named it in an answer.

On the other 39 questions no run cited or named it, on any engine, in either wave: 2,340 runs in wave 2 and 1,340 in wave 1. The question that asks about AI visibility is a probe, and nothing found only on it is a finding about the Dubai market.

05Bounds

What this study does not claim

The overlap figure measures whether two engines name the same firms for the same question. It does not say whether either engine names good firms, it does not say which engine is right, and it says nothing about why engines agree or differ. We had no view into how any of them retrieves or ranks.

A firm appearing in more answers is not a firm recommended more strongly, and the tables are not a league table. A firm missing from an answer was not named in that answer, and that is all its absence means.

The table of the 20 firms named in the most runs, below the findings at the top of the page, is ordered by those runs and is not a league table either: being named in more runs is not a quality ranking and not an endorsement.

Low overlap between engines does not mean their answers are random. Each engine agrees with itself, run to run, more than it agrees with the others, and when the engine label is shuffled that pattern goes: 27.6% of pairs fall below the mean on average, against 90.4% with the real labels. What was measured is that the engines are each consistent and differ from one another, not that any of them is arbitrary.

An unresolved name is not an unlicensed firm. A citation is not an endorsement.

06FAQ

Questions buyers ask about this study

07Review

How the rules and the code were reviewed

The code that applies the extraction rules went through four review rounds, on 26 September and 1, 2 and 4 October. Each round read the rule text and the code against each other and reported what departed from it.

All four started with fresh context. What protects their findings is what the reviewers had seen, and they were not run the same way:

• Round 1 was a separate Claude Code session that the study's owner opened on 26 September, with fresh context. Its record states what it reviewed, not what it opened. It found three places where the code departed from the rules and five where the rules left the decision to the code. Each was decided in writing before any re-extraction.
• Rounds 2, 3 and 4, on 1, 2 and 4 October, were subagents spawned by the session that wrote the code, working from briefs that session wrote. They had no access to its conversation. The builder chose what each one was told, including what it could not open. Each records that it ran no wave 2 analysis script and opened no wave 2 output; round 4 adds that it ran no gate, named-set or overlap computation.

The findings of all four rounds ship in the bundle.

08Limitations

Limitations

This edition measures which real estate firms three AI search products name in response to 40 buyer questions about Dubai, 20 runs per question per engine, collected on 23 September 2026. The measurement basis, register-resolved firms in text channels across every run of a cell, was committed to git before the data was inspected. What follows is what that design cannot see, and what we found while checking it.

What the figures count

Every figure counts a firm as named when a register-resolved firm name appears in the response's text channels. Two consequences follow.

A firm appearing only in a listing block is not counted. These blocks carry firm names with postal addresses, sourced from local business listing data rather than written by the engine. A firm appears there because it maintains a business profile, which is a different claim from an engine recommending it.

They are excluded by the committed basis, and the exclusion is not evenly spread: Gemini answers carry no listing block in either wave (0 of 300 runs, then 0 of 800); ChatGPT answers carry one in 88.0% of runs in wave 1 and 89.8% in wave 2; Google AI Mode in 54.0% and 50.7%, from 0 of 20 runs on some questions to 20 of 20 on others (analysis/dubai-02-wave2/card-rates.txt).

A firm whose name does not resolve to the Dubai Land Department register is not counted. Unresolved strings are recorded and reported as a count, not as firms. That includes a firm named only by an abbreviation no register name contains: "JLL" does not resolve to Jones Lang LaSalle's registered brokerage, and is not counted for it.

Resolution can credit a mention to a different registered firm than the one the text names: under R2, a listing card's contact number can point from the name in the text to another firm. In wave 2, 2 of the 188 counted firms are credited this way and are never named by their own name in any run (analysis/dubai-02-wave2/r2-sensitivity.txt).

One of them is a case R2 named before collection: the rule committed on 19 September cites Christie's on dubai-01's evidence. In wave 2, answers naming Christie's International Real Estate are credited to Premier Estates, which makes it the 13th most-named firm in this edition, at 229 runs across 45 cells on all three engines.

Christie's own registered entity has had no active broker card since July 2016. The contact number on the listing cards beside those answers is registered to Premier Estates' brokers, which is R2's first test. The register records no other link between the two, and it does not say Premier Estates holds the licence for the Christie's operation.

The response shape the figures sit on

Responses are not uniform. There are at least five layouts, and which one appears depends on the engine and on the question. What separates a firm's name from its description, where links sit, and whether the response carries a structured listing block all change between layouts.

Where a response carries both a listing block and prose, most firms in the block are never discussed in the prose. In the runs examined by hand, two of nineteen firms appeared in both.

Responses commonly end by asking the reader to state a budget or an area before narrowing further. Every figure here comes from the first turn of an interaction the engine itself treats as unfinished. A reader who answers the follow-up gets a different and narrower list, and this study has no visibility into it.

Google AI Mode's listing counts cluster hard at exactly 20 per response, with a tail beyond. That pattern is consistent with a page size rather than a count of what exists, so AI Mode listing counts should be read as page-limited.

Name resolution

The same firm is written several ways inside a single response, including differences of case, accent and apostrophe. Counting every unresolved string that carries all the distinctive words of a firm's registered name, or that the name carries, as naming that firm moves wave 2's share below the mean, by this test, at most 1.7 percentage points (90.4% to 88.7%) and its median ratio by 0.010; abbreviations and trade names it cannot see are outside it.

On wave 1 it moves the share by 4.8 percentage points (88.1% to 83.3%) and its median ratio by 0.016. Both are differences between two shares, so the two waves' bounds are in the same unit.

Those bounds do not reach firms whose names contain no word of three letters or more, and those firms are undercounted by an amount we have not measured.

On Google AI Mode, an answer card is credited to a registered firm when its opening two or more words match that firm, with no ruling on the individual card: 132 card names in wave 2, in 171 mentions across 24 firms, and 24 in wave 1, in 29 mentions across 10 firms. Each was read by eye and the four in doubt were ruled; the others depend on that reading, not on a rule the code applies.

A dormant register entry resolves like an active one, because the basis is registration, not activity. On the pre-registered basis, 4 of the 178 firms resolved in wave 1 and 3 of the 188 in wave 2 are dormant in the register, named in 7 and 10 cell counts; 1 and 2 more hold a dormant entry beside an active one (dormant_resolved.py).

Firm names built from surnames are read as people by matchers and by human reviewers alike. One such name was ruled a person during review and corrected on a second pass.

Person detection, and what it cannot see

The study reports where individuals are named. The detector that finds them has two known limits.

It cannot propose a name of a single token, because matching one token against a register of full names floods. It cannot see a name glued to the word that follows it.

Five register-held individuals were hidden this way. Two reached review only because an independent check happened to list them.

One was proposed without the surname that was glued to the next word, so its ruling removes the rest of the name and leaves the surname. Two were surfaced by neither the detector nor that check, and were found only when the glued-name output was examined directly.

To bound the second, we scanned the answer layer for capitalised sequences and removed everything accounted for by the candidate list, the resolved firms, the register, and lists of places, months, languages and interface text. 828 distinct sequences remain unaccounted for, at the person rulings of 4 October, when the scan ran. The rulings made on 5 October strip two more, so the same scan on the rulings of 5 October and since leaves 826; neither of the two was in the sample below, which was drawn from the 828. A random sample of 50 of them (seed 20261004) was classified by hand: 1 person name, 12 firms, 13 places and 24 other. That gives a point estimate of 16.6 person-name sequences among the 828, with a 95% Clopper-Pearson interval of 0.4 to 88.2 (rate 0.05% to 10.65%; Wilson 2.9 to 86.9).

The interval counts sequences, and one person can appear in several, so the number of individuals it implies is no larger. Computed by analysis/dubai-02-wave2/miss_scan_bound.py. Names written as one word are not bounded: the scan starts from runs of two or more capitalised words.

Every person count in this edition is therefore a floor, and this is the bound on how far below the true count it may sit for names of two or more words in both waves' answer layer, the layer the scan read. 22 of wave 2's 154 people come only from link addresses and listing cards, outside that layer. The scan itself was validated by that sample: it surfaced a person the detector had never proposed, a name with a job title run on after it, which a synthetic control had not shown.

Where named individuals come from differs by engine, and the difference is larger than the aggregate shows. Of the 154 individuals ruled and redacted who appear anywhere in wave 2's responses, 66 first appear in the engine's own answer text and 66 only in text quoted from a cited page. The other 22 are 4 found in a link's address and in no answer text, and 18 found only in what the engine presents around the answer, such as listing cards.

Underneath that tie the three engines behave differently. Google AI Mode volunteered an individual in its own prose once in 32, with 28 arriving from cited pages.

ChatGPT did close to the reverse: 2 of its 52 came from cited text, and the rest from its own prose or from the listing cards around the answer. Gemini sits between them, with cited pages ahead in both waves. Wave 1, with fewer runs on two of the three engines, shows the same shape.

A name reaching the response through a cited page was written by that page, not by the engine. On this evidence, the claim that these systems know individual Dubai agents holds very differently depending on which system is asked. The population is every person ruled and redacted, searched in each wave's responses and grouped into individuals, a person counted once however many forms of the name appear (provenance.py).

What was read by a human, and what was not

A human read 20 responses cold, counting individuals by eye, before seeing what the detector had found. That read covered 2 of the 40 questions, because the extract it used defaulted to two question ids.

Both of those questions ask which firm to use. Other questions in the pack do ask about brokers and agents, and responses to them do name individuals.

The file used for that read had engine identifiers removed, but the response layouts differ enough by engine that a reader who knew the layouts could recover the engine. The reader did not know them at the time. The read is reported as engine-stripped rather than blind.

During the later review, a bulk action ruled 104 rows without displaying them, so 104 evidence passages went unread. The missed-name log gathered during that review is a floor on what a reader would have found.

Pre-registration

The replication thresholds and the measurement basis were committed at 17e8485, 30 minutes after the collection run began and before any extraction ran. No result existed when the commit was made, and the commit is therefore described as made before the numbers existed rather than before the run started.

The times of the collection and the commit are published on this page so this can be checked. The first extraction's exact time was not recorded: it ran after the last request, between 05:41:21 and 10:47 UTC on 23 September. Times here, as everywhere on this page, are UTC.

While preparing the extraction rules, the session writing them printed wave 2's headline figures, with no rulings applied, between 20:58 and 21:07 UTC on 25 September, before it wrote the implementation (1b65dce). That was disclosed at the time and recorded in the committed rules.

The implementation was then reviewed against the rule text by a separate session opened with fresh context on 26 September; its record states what it reviewed, not what it opened. The review found three defects: two removed firm credit and one raised AI Mode's distinct-string count. All three were fixed on 29 September (e9984f7), after the 25 September figures and before the statistic's first recorded run on the rules this page uses, at 13:48:06 UTC on 5 October.

The 23 September gate.py run and its timing relative to extraction are described above, under "A run that wrote the statistic's inputs".

Things that change on their own

Responses carry values that move without the recommendation moving: star ratings, review counts, opening status, listing totals and quarterly awards. None of them enters the replication statistic, which compares the sets of firms named. They are excluded from every wave-to-wave comparison here, because a difference in them measures the world changing rather than the engine changing.

Citations

A citation in these responses is not evidence for the claim it sits beside. Firms are cited to competitors' blogs, to unrelated company directories, and to opaque search-viewer URLs with no readable source. Nothing in this study treats being cited as a quality signal, and no reader should.

Citation counts are not reported in this edition. Cited URLs ship as their host and a hash of the rest.

09Corrections

Corrections

This log is permanent and carries into every edition. It holds what was wrong before publication as well as after.

• The measurement basis gained a channel after registration. answer_card was added to the naming channels on 26 September (1b65dce), three days after collection, and recognised as a change to the registered basis on 4 October before it reached any figure. The headline is computed on the registered pair; the extended basis is reported beside it without a verdict.
• The implementation of the extraction rules departed from the rule text in three places. Review round 1 found them on 26 September. Two of the three removed firm credit. All three were fixed in e9984f7, committed at 07:52:46 UTC on 29 September, six days before the statistic's first recorded run on the rules this page uses, at 13:48:06 UTC on 5 October. They were not fixed before every wave 2 figure: headline figures were printed on 25 September, before the review, as the pre-registration section discloses.
• gate.py, which builds the per-cell tables the statistic reads, resolved the first form it saw of each name and applied that result to every other form of it, so 13 mentions in 13 wave 2 runs and 4 in wave 1 took another form's result. The independent recomputation found it on 6 October. It was fixed (31817ee) and every figure recomputed: both headline figures and the verdict are unchanged, and question F4's row in the per-pair table moved.
• A firm whose name is built from surnames was ruled a person during review, and corrected on a second pass.
• A person was proposed for review without their surname, because the surname was glued to the next word. The ruling covers the rest of the name and leaves the surname. Recorded under Limitations.
• Wave 1's resolution carries corrections made on 21 September after its output was read, each disclosed as a post-hoc correction. The general rule that would have caught them, trade-name matching, was committed for wave 2 before wave 2 was collected. It is not implemented in this edition and is carried to the next. The rule has three tests. Containment is bounded: 3 unresolved strings per wave sit inside exactly one active registered name and would be resolved by it, 1 of them at the publication floor in wave 2 and none in wave 1. Register corroboration and own-site declaration are not bounded in this edition.
• R5, which lets a listing card's contact number alone resolve a firm that no name reached, was committed for wave 2 on 21 September, before wave 2 was collected. It is not implemented in this edition and is carried to the next. No figure on this page depends on it, for the reason given under The register.
• The disclosure of the 25 September print in committed-rules.md, written on 26 September, says the figures "were not repeated or used". That says more than the record establishes: the record shows that no commit holds the figures and that no figure on this page is computed from that run, not that they had no influence on R6 and R7. The page's narrower account, under "A rule on the pre-registered channels, written after a figure had been seen" and "A session that saw figures", is the one to rely on. The original sentence is left as committed, and a note dated 12 October sits directly below it in rules/committed-rules.md in the bundle.
• Names on a check list were published unchecked, and two place names were credited as firms. A list of strings resolved to a licence with no active broker, "each to be checked before it is published", held eleven entries. Four were checked before publication: two spellings of Christie's International Real Estate, Powerhouse and Totality Commercial. Seven were not. Found on 13 October 2026, before this page was republished, and ruled under two rules committed and pushed before their effects were computed (4d4cecc, d20c50e):
  • "Arjan", a Dubai district, was credited to ARJAN REAL ESTATE BROKER in 18 wave 2 runs and 4 wave 1 runs. All 18 wave 2 runs are Google AI Mode answers to C1p, where it sits in a table headed "Community", between Al Furjan and Dubai South, with no listing card, link or contact number behind it. Removed.
  • "Jumeirah Golf Estates", a community, in 2 runs. Removed, as a decision made before wave 1 and never applied required.
  • Bare "Vacationer", in 1 run, named among holiday-home operators. Classed with Vacationer Holiday Homes, a holiday-homes operator and not a broker.
  • "URBAN", in 1 run. One common word that begins 46 registered names, credited to a licence with no broker since 2016 while the answer cites live listings. Now unresolved.
  • "Blueground" and "Fine & Country", in two spellings, 3 runs each. Checked and kept: each is the registered name of the firm the answers name. Both licences are dormant and are counted as dormant.
  Checking the study's own list of Dubai place names against the credited firms found one more: "Dubai Marina", credited to an active licence in 3 runs of each wave. Removed. Wave 2's headline is unchanged at 104 of 115 pairs below the mean (90.4%), median ratio 0.652, and so are the verdict, the figure on the extended basis and every drift verdict. Wave 2 counts 188 firms rather than 193. Wave 1's median ratio after re-extraction moves from 0.676 to 0.656, its share unchanged at 37 of 42. The figures derived from these move with them, each corrected where it appears: the wave 1 comparisons and its spelling bound, the shuffled baseline (27.9% to 27.6%), the spelling counts, the dormant-entry and unresolved counts, and the tables of eight questions.
• A committed rule asserted a check that had not been performed. The rulings of 5 October say that a hand class gives way to the register, and that none of the 13 other hand-classed brands has a registered firm. Three do, each only as a dormant entry: Sobha, DAMAC and Vacationer. Which 13 the sentence meant cannot be told from the text or the history, so on 13 October 2026 every brand in the four classes that name a business, 30 in all, was checked against the register. Four holiday-homes strings also name a brand the register holds, and none of them resolves to it with its class removed. The three keep their classes under the reading set out with the hand classes, which changes no published figure, and the figures under the reading not taken are published with the other checks. The original sentence is left as committed, with a note dated 13 October below it in rules/committed-rules.md in the bundle. The same day, the count of 828 sequences behind the bound on missed names was dated to the person rulings it was computed on; on later rulings it is 826.

10The bundle

Check this yourself

The evidence bundle is at /research/dubai-02/data/, with its manifest readable unzipped at manifest.txt. It holds the question pack, the redacted ledger tables, the per-cell tables the headline and the firm tables were computed from, the scripts, the commit and push timestamps above, the four review records, and, in each wave's runs.csv, the SHA-256 of every stored answer.

The headline reruns from the bundle on the Python standard library alone: python3 analysis/dubai-02-wave2/headline.py wave2/cells/q1_surface_overlap.csv returns 104 of 115 pairs below the mean (90.4%), median ratio 0.652, and the verdict REPLICATES.

The cell tables themselves cannot be regenerated from the bundle. That takes the Dubai Land Department register, which is third-party data we are not free to redistribute, and the raw answers, which are not shipped because they carry the names of individuals we ruled must not be published. Both are decisions, not gaps.

The other scripts ship so the method can be read, not so it can be run: among them gate.py, which writes the cell tables, and recompute.py, the independent reimplementation.

It does not hold the answers' raw text. The detector that finds individuals' names has limits, set out above, and a name it cannot see in raw text would be published if raw text shipped.

The hashes let anyone who is given an answer check that it is the one measured, and we send the raw answers privately to anyone who asks. dubai-01 shipped its raw answers; dubai-02 does not, and this is why.

Every file in the bundle passes the same phone and name scan as the repository before release.

The bundle's code, under analysis/, is licensed under the Apache License 2.0, and everything else in it under Creative Commons Attribution 4.0 International (CC BY 4.0), as its LICENSE and LICENSE-DATA files state. CC BY means anyone may reuse and adapt the data, provided they credit it: Aaron Zara, dubai-02 evidence bundle, RealEstateSEO.ph. The licence covers the rights the author holds; it grants none in the engines' text or the third-party listings the tables reproduce.

11Sources

Sources

• Dubai Land Department broker register, through Dubai Pulse: dld_brokers-open, snapshot of 16 September 2026. A later download will not match the published hashes.
• DataForSEO API documentation for the three endpoints used: ChatGPT, Gemini, Google AI Mode.
• The pre-registered procedure and threshold, wave2-decision.md at 17e8485, and the extraction rules, committed-rules.md, both in the bundle.
• dubai-01, the single-engine pilot this study follows: /research/dubai-01/.

AAppendix

Appendix: the 40 questions, verbatim

Every question was asked on all three engines, 20 times each. G4 names no firm by design and sits in no verdict.