Research conducted by Lunasol. On 21 September 2026 we put one question to Google's AI Mode, "Which are the best [sector] in Dubai?", and ran it across twelve kinds of supplier. The twelve answers cited 109 host slots between them. Eighty of those, 73 per cent, were the website of a business that sells the thing being asked about. This article reports what came back, in full, with every figure set against the sample it was drawn from. Lunasol sells work aimed at how businesses are described in these answers, which is why the protocol and the classification scheme were both written and frozen before the first question was asked.
AI Visibility is owned by Lunasol.
The short version: that 73 per cent, 80 of the 109 cited host slots, is the largest category by a wide margin, and the remaining 29 of the 109 are spread across five categories. Three categories we had defined in advance, before seeing any data, came back empty. No platform documentation, no government or official body and no education or research source appeared in any of the twelve answers. The mix moves with the sector, and the sector with the lowest supplier share was also the only one where the answer attributed its list to independent directories. Our guess is that the sectors with an established independent ranking body are the ones that lean on suppliers least, but we did not record which sectors have one, so that is a hypothesis and not a finding. It is a pilot: one surface, Google's AI Mode, on one day, twelve runs. It says nothing about any other assistant. Everything below states its denominator.
What we did
Twelve sectors, one question shape, one run each, on one surface, on one day. A classification scheme written after the data arrives is a scheme written to suit the data. Fixing the scheme in advance does not make a small sample any bigger, and it is no substitute for one. What it does mean is that the categories were not chosen to fit the numbers. Lunasol froze this protocol, and the scheme with it, before the first run.
One question shape, twelve sectors. Every run used the same sentence with one word changed: "Which are the best [sector] in Dubai?" The twelve sectors, fixed in this order before any run: interior fit-out companies, real estate agencies, business setup consultants, restaurants, accounting firms, IT support companies, law firms, dental clinics, gyms, nurseries, car service centres, moving companies. Because the shape never changed, no difference between the answers is attributable to the wording. With one run per sector we cannot separate a sector effect from ordinary run-to-run variation.
One surface, one run each, one day. Google's AI Mode, signed in, one browser, one connection, all twelve on the same day. We did not attempt a second assistant, and that was decided in the protocol before the runs rather than after them: the second assistant this series has tried has returned no usable answer from the browser, account and location we work from, across three earlier articles, the cause was never established, and we do not report a failure we cannot explain. So this is a study of one surface.
Nine categories, fixed in advance. Every cited host was placed in exactly one: the website of a business that supplies the thing asked about; a directory or listing platform; news or trade media; a user content platform such as a forum, a social network or a video platform; a product's own documentation; a government or official body; marketing or vendor guidance from a supplier in an adjacent field; education or research; and other. The tie-break rule was also fixed in advance: where a host could sit in two categories, what the host primarily is beats what the cited page happens to be about.
What was counted. For every run: the total outbound links that were not Google's own, the distinct hosts among them, the category of each host, how many businesses the answer named, whether it produced a comparison table, whether it produced a map card, and whether it added any disclaimer beyond the standard footer. By map card we mean the block of local results with a map in it. By the standard footer we mean the disclaimer the surface appends itself, which we did not count as an addition by the answer. Counts were read from the rendered page. We did not open a single cited page. So everything here is a statement about what was cited, not about what those pages say.
No failures. Twelve runs, twelve answers, no error, no truncation, no empty response, so no re-runs were needed.
One counting rule, because it affects a number below. The www and non-www forms of the same domain were counted as two hosts. That rule was fixed before the data was seen, and changing it afterwards would be fitting the rule to the result. It happened twice, both times in the same run, so that run's fourteen hosts are twelve organisations, and both figures are given. Under the other rule the study total would be 107 host slots rather than 109. We did not record which of the nine categories those two duplicated hosts fell into, so the supplier count would be 78, 79 or 80 out of 107, which is between 72.9 and 74.8 per cent. Either way supplier websites stay the largest category by a wide margin, though the rounded share would read 73, 74 or 75 per cent instead of 73.
What did the twelve answers cite? The headline number
Lunasol put one question, "Which are the best [sector] in Dubai?", to Google's AI Mode across twelve sectors on 21 September 2026, one run each. The twelve answers cited 109 host slots, and 80 of them, 73 per cent, were the website of a business that sells the thing being asked about. A host cited in two sectors counts once in each, so 109 is a per-run count and not a count of distinct organisations across the study.
Everything else, all eight remaining categories together, came to 29 host slots out of 109. The table below accounts for all 109.
| Category | Host slots | Share of 109 |
|---|---|---|
| Website of a supplier of the thing asked about | 80 | 73 per cent |
| Directory or listing platform | 11 | 10 per cent |
| User content platform (forum, social, video) | 8 | 7 per cent |
| Marketing or vendor guidance from an adjacent supplier | 5 | 5 per cent |
| News or trade media | 3 | 3 per cent |
| Other | 2 | 2 per cent |
| Product or platform documentation | 0 | 0 per cent |
| Government or official body | 0 | 0 per cent |
| Education or research | 0 | 0 per cent |
In one sector of the twelve, car service centres, every cited host was a supplier website. Eight hosts, eight supplier sites, no directory, no media, nothing else. That is one run.
The twelve runs in full
Here is every run, as Lunasol recorded it on the day. Links are the total outbound non-Google links counted on the rendered page. Hosts are the distinct websites those links point to, counted once each per run, so a website cited four times in one answer is one host. Supplier sites are the hosts that belonged to a business selling the thing asked about, and Everything else breaks the rest down by category.
The last three columns are observation rather than classification: how many businesses the answer listed, and whether it produced a comparison table and a map card. Businesses named counts businesses, not sources, and no category share in this article is calculated from it.
| Sector | Links | Hosts | Supplier sites | Everything else | Businesses named | Comparison table | Map card |
|---|---|---|---|---|---|---|---|
| Interior fit-out companies | 24 | 8 | 7 | 1 user content | 9 | yes | no |
| Real estate agencies | 27 | 13 | 8 | 2 user content, 2 adjacent vendor, 1 other | 7 | yes | yes |
| Business setup consultants | 15 | 7 | 6 | 1 user content | 6 | yes | yes |
| Restaurants | 16 | 9 | 8 | 1 directory | 9 | no | yes |
| Accounting firms | 20 | 8 | 6 | 1 directory, 1 adjacent vendor | 7 | yes | yes |
| IT support companies | 31 | 14 | 9 | 4 directory, 1 user content | 8 | yes | yes |
| Law firms | 23 | 8 | 4 | 2 directory, 1 media, 1 user content | 12 | yes | yes |
| Dental clinics | 12 | 6 | 5 | 1 adjacent vendor | 8 | yes | yes |
| Gyms | 15 | 7 | 5 | 1 media, 1 user content | 6 | yes | yes |
| Nurseries | 16 | 10 | 6 | 3 directory, 1 other | 7 | yes | yes |
| Car service centres | 16 | 8 | 8 | none | 8 | yes | yes |
| Moving companies | 20 | 11 | 8 | 1 media, 1 user content, 1 adjacent vendor | 12 | yes | yes |
| All twelve runs | 235 | 109 | 80 | 29 | not totalled | 11 of 12 | 11 of 12 |
Links per run ranged from 12 to 31, with a median of 18. Hosts per run ranged from 6 to 14, with a median of 8.
The IT support run is the one affected by the www counting rule: its fourteen hosts are twelve organisations, because two hosts were each counted under both forms of their domain.
The businesses-named column is raw data and is published here for completeness. It was not all counted the same way: in ten runs it was read from the answer entry by entry, and in the last two it was derived by matching the answer's own entry pattern. Two of those counts need a qualifier the table does not carry: in the accounting run seven businesses were named individually and four more named only as a category, and in the nurseries run seven were named in the childcare sense and two were plant nurseries, because the answer covered both meanings of the word. The relationship between being named in an answer and being cited as a source in it is a separate question. It is the subject of our later article on the difference between a mention, a citation and a recommendation, which ran its own set of runs on these same twelve sectors, because this study recorded how many businesses were named and not which ones. Both articles say where their data came from.
Three categories that never appeared
The scheme had nine categories. Three of them finished these twelve Google AI Mode runs empty, and because they were defined before any data existed, three zeros are a result rather than an oversight.
No product or platform documentation. Not one of the twelve answers cited a search engine's or an assistant maker's own documentation about how any of this works. That is worth holding next to what those companies publish. Google states that "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary" and that "You don't need to create new machine readable files, AI text files, or markup to appear in these features" (Google Search Central). Those statements exist, they are public, and on this sample they were never the thing an answer leaned on when recommending a supplier.
No government or official body. Across the 109 host slots, no regulator, registry or official body was cited once. In one run the answer told the reader to check a public register, without citing the register or anyone who keeps one.
No education or research source. No education or research source appeared in any of the 109 host slots across the twelve runs. No university and no research body was cited in any of them.
We are not going to explain why. We did not measure why, we know of no assistant that publishes the reason a source is chosen, and a reason we made up would be worth nothing to you. What we can say is what was there instead, and 80 of the 109 host slots, 73 per cent, were the suppliers themselves.
Does the source mix change between sectors?
It does, and this pilot cannot say what drives it. The supplier share runs from eight of eight cited hosts in one sector to four of eight in another, and with one run per sector we cannot separate a sector effect from ordinary run-to-run variation.
Of the twelve sectors we ran, law had the lowest supplier share, at four of its eight cited hosts. It was also the only sector where the answer attributed its list to independent directories, and two of its eight cited hosts were directory platforms. One sector is one sector.
IT support carried the most directory host slots of any run, four of its fourteen, and those fourteen host slots came from twelve organisations. That run was also the only one to carry pricing, attributed to the directories rather than to any supplier.
Four sectors ran at six in seven or better on supplier sites. Car service centres at eight of eight, restaurants at eight of nine, interior fit-out at seven of eight, business setup at six of seven. We did not record whether a sector has an independent ranking body, so we cannot say that is what separates them, and we did not check what neutral rankings exist in any of these sectors.
That is an observation about twelve runs on one day, not a law. Read the ranking-body idea as a hypothesis worth testing on a larger sample.
Seven things we did not go looking for
Alongside the counts, Lunasol's record of the runs carried observations the classification scheme does not capture. These are the seven that stood out.
One, suppliers are cited as the authority on rankings of their own market. In the fit-out run, three of the seven supplier sites were cited for pages ranking the best companies in their own category. In the accounting run, the most cited source on who the best accountants are was a blog post by a software vendor in an adjacent field. In both runs the ranking came from inside the market, or from beside it.
Two, in one run the authority invoked was not the source cited. In the restaurants run the answer rested on the recognition of one international restaurant award body in five of its nine entries. No host belonging to that body was cited anywhere in that run. The claim arrived through a travel booking site and through the restaurants' own pages.
Three, the extra disclaimers are inconsistent. The accounting answer added "For financial advice, consult a professional" after the standard footer. The dental answer added "This is for informational purposes only. For medical advice or diagnosis, consult a professional." before it. The law answer, on a regulated profession, added nothing at all. The gyms answer added nothing either.
Four, individual people are named and praised. The real estate answer named six individual agents and attributed client praise to each by name. The dental answer named ten individual doctors and attributed praise for specific clinical procedures to several of them. These are private individuals, named in a machine-generated answer, with quality judgements attached to them.
Five, one sector of twelve carried prices, and they came from a directory. The IT support answer published hourly bands and a project minimum inside its comparison table, attributed to directory platforms rather than to any supplier. No other run carried a price of any kind. We are not reproducing the figures, because a figure with nothing neutral behind it becomes an anchor the moment a reader sees it.
Six, the answer sometimes argues against the category it is describing. The business setup answer closed by telling the reader they could apply directly to a free zone authority and bypass consulting fees altogether. The accounting answer told the reader to verify an auditor against a public register. Neither was asked for. We record what the answers said and we did not check either statement, so neither is our advice about what any UAE process requires.
Seven, operational detail is published that nobody asked for. The business setup answer published street addresses down to the office number for all six firms it named. The accounting answer published opening hours for each firm. Neither was asked for, and we did not check either against what those businesses publish themselves.
There is an eighth that is not a finding so much as a shape. Eleven of twelve runs produced a comparison table. Eleven of twelve produced a map card. Twelve of twelve ended by asking the reader a question and offering numbered options. The format is close to fixed.
What this pilot cannot tell you
Read this before you quote any number above.
Twelve runs is twelve runs. One question per sector, one sector per run. Every figure here rests on a sample you can count on your fingers twice, and none of the shares should be treated as stable. We call it a pilot for that reason.
One surface, one day. Everything here describes Google's AI Mode on 21 September 2026. It says nothing about any other assistant, and nothing about any other day.
One question shape. "Which are the best [sector] in Dubai?" is a category question asked by somebody with no supplier in mind. A buyer who phrases the question differently is asking a different question. Whether that changes the source mix is the subject of our later study on what changes when a question gets more specific, and this one cannot answer it.
One run each, so nothing here is about variability. Whether the same question run twice returns the same source mix is the subject of our later study on variability across repeated runs, and this one cannot speak to it. Nothing above should be read as saying the mix is stable between runs.
The classification is our judgement. Nine categories and a tie-break rule, both fixed in advance, applied by us. Another classifier would place some hosts differently, particularly the adjacent-vendor and other categories, which are the two most arguable, and which hold five and two of the 109 host slots between them. A reclassification of those seven hosts would not move the headline, but it would move the small numbers.
We did not open any cited page. So this is a record of what was cited, not an assessment of whether any of it was any good.
It does not tell you why. We know of no assistant that publishes the reason a source was chosen. Correlation between what exists and what got cited is not evidence of a mechanism.
And it does not promise anything. Nothing here makes a business cited, named or recommended. Google states that "No one can guarantee a #1 ranking on Google" (Google Search Central), and OpenAI states that "Placement is not guaranteed" and that "Search results and citations can be incomplete, outdated, or incorrect" (OpenAI Help Center). Our article on whether anyone can guarantee a ChatGPT recommendation sets out the whole position.
What a UAE business can do with this
This is the only section where a finding turns into advice. It is general information about what our twelve runs cited. It is not legal, financial, medical, regulatory or other professional advice. It does not tell you anything about licensing, or about what any regulated sector requires. If you work in a regulated sector, take your own professional advice before you act on any of it.
Supplier websites were the largest category these answers cited, and your own site is that kind of site. On this sample, in this sector spread, on this surface, on this day. Being that kind of site is not the same as being cited, and nothing here says yours will be. What it means is that on the sample we took, the thing cited most often was the sort of thing you control. The three things you cannot control (regulators, documentation and research) were cited not at all. It costs nothing to check which page on your own site would answer the question a customer actually asks, or whether that page exists at all. In our fit-out run, three of the seven supplier sites were cited for pages ranking the best companies in their own category. That is three hosts in one run on one day, and it does not tell you which page of yours would be cited, or whether any would.
Find out whether your sector has an independent ranking body, and whether you are in it. In one sector of twelve, the answer attributed its list to independent directories, and that run had the lowest supplier share of the twelve. That is one run, and we did not record which sectors have such directories at all. If your profession has a recognised directory, whether you are in it is a specific, checkable fact. Being in it does not make you cited, and nothing in our data says it does. If your profession has no such directory, that part of the map is empty for everyone.
Read an answer's citations separately from its claims. Our restaurants run invoked an award body in five of its nine entries and cited no host belonging to it.
If an answer names a member of your staff, you should know that before a customer does. Two of our twelve runs named individuals and attached quality judgements to them. That is a thing to monitor deliberately rather than to discover.
Do not build a plan on twelve runs, including ours. Run the same shape of check on your own sector first. Our article on how to check what AI says about your company sets out how to do that.
Each of those is a check you can run on your own market, with your own questions, at whatever depth you want.
How this study relates to Lunasol's published process
Lunasol's published process begins with an audit, described on our service page (Lunasol) as "The AI census". The same page lists "Presence in sources AI models trust and quote" and "AI visibility audit, what ChatGPT says about you today" among the GEO deliverables, GEO meaning work aimed at how a business is described in AI answers.
The published audit is described as "what ChatGPT says about you today": one business, on the assistant that line names. This article asks a different question on a different surface: which categories of source were cited across twelve sectors in Google's AI Mode. This study speaks only for the surface it ran on. The published process ends with a step called Track, described as "Numbers, every quarter", and what an audit contains is set out in our article on what an AI search audit should include.
What to do next. Lunasol's GEO and AEO scope is published at lunasol.ae/geo-aeo, and our article on what Lunasol's service covers goes through it in detail. If that is the reading you want for your own business, ask for the AI visibility audit.
Common questions
How many runs was this? Twelve. One per sector, on one surface, on one day. Every study-wide share here is a share of the 109 cited host slots those twelve answers produced, except the alternative figure at the counting rule, which is stated out of 107. Every other figure names its own denominator where it appears, whether that is one run's hosts, one run's supplier sites, the entries an answer listed, or the twelve runs themselves.
Why only one assistant? Because this environment produced no usable answer from the second assistant this series has tried, across three earlier articles. The cause was never established. We make no claim about that assistant itself, and we decided before these runs rather than after them that we would not report a failure we could not explain.
Does 73 per cent mean my website is 73 per cent of the answer? No. It means that of the hosts cited across our twelve answers, 73 per cent were supplier websites in general, not yours. Being in the category that gets cited is not the same as being cited.
Why did you publish the empty categories? Because we wrote the categories down before we had the data. A scheme that only reports the boxes that filled up is a scheme being edited after the fact.
Can I see the raw data? The per-run record exists, with every host and its classification. It carries company and domain names, which we do not publish. What this article gives you is that record with the names removed and the counts kept, which is the per-run table above. There is no separate data file and no separate published report anywhere.
Are you going to run this bigger? The useful next step is more runs per sector and more than one surface, which is a different piece of work. The design above is the one a larger version would start from.
Method, sources and notes
Research conducted by Lunasol. Data collected 21 September 2026 in Google's AI Mode, signed in, one browser, one connection, one day, under a protocol written and frozen before the first run. Twelve sectors, one prompt shape, one run each. No failures and no re-runs. Counts were read from the rendered page; no cited page was opened. The classification scheme and its tie-break rule were fixed before any data was seen, and three of its nine categories finished empty. The arithmetic was checked: the twelve per-run classifications sum to 109, which matches the host-slot total.
One method difference, reported rather than smoothed. In ten of the twelve runs the count of businesses named was read from the answer entry by entry. In the last two it was derived by matching the answer's own entry pattern in the rendered text. The two methods are not identical, they affect only the businesses-named column, no category share in this article is calculated from that column, and the two figures stated against it, the restaurants entry count and the business setup firm count, both come from runs that were counted entry by entry.
The related study. Our later article on the difference between a mention, a citation and a recommendation examines that difference. It does not use this dataset. This study recorded the number of businesses an answer named and not their identities, so matching a named business to a cited host could not be done from it. That article ran its own set of runs on these same twelve sectors, on the same day and the same surface, under its own protocol frozen before its first counted run. Two studies on the same twelve sectors are not two independent confirmations of each other, and both articles say so.
Names. No host, domain, company, agency, publication or individual from the study is named anywhere in this article. Sector labels are categories, not businesses. The two answers that named individual people are reported by count and by category of claim, and the names are not reproduced here in any form.
Who publishes this. AI Visibility is owned by Lunasol, a Dubai agency that sells work aimed at how businesses are described in these answers, which means we have a commercial interest in one of the findings above. The protocol was frozen before the data existed precisely because of that, the classification scheme with it, and the section where a finding turns into advice is separated and labelled. The author writes for this publication, which Lunasol owns, and is not independent of it. This article makes no claim about any other provider, named or unnamed, in any direction.
This is a pilot. It is general information and not legal, financial, medical or professional advice of any kind. Nothing here makes a business shown, named, cited or recommended in any AI product, and nobody can promise that it will. Assistant answers can vary by account, location and time, and quoted documentation changes without notice, so check the current version of any page quoted here before relying on it.
Product and company names are used here for identification only. Neither this publication nor its owner is affiliated with, endorsed by or sponsored by any of them, nothing here is an endorsement by us of any of them, and none of them has reviewed, approved or contributed to this article. ChatGPT is a product of OpenAI, and the OpenAI Help Center is a publication of OpenAI. Google Search, Google Search Central, AI Mode and AI Overviews are products or publications of Google. All are trademarks of their respective owners.
AI Visibility is owned by Lunasol.
