Does a More Specific Question Change Which Sources Google AI Mode Cites? A Lunasol Study

Research conducted by Lunasol. In the twelve answers we recorded, the more detailed questions were answered with different sources, and not in proportion to the detail added. Of the 31 websites the plain question cited across four sectors in Google's AI Mode on 21 September 2026, 17, or 55 per cent, were still there when one qualifier was added, and 4, or 13 per cent, were still there in the most detailed version, against the 74 to 77 per cent that the identical question kept when we asked it again with nothing changed. That 74 to 77 per cent is not measured here. It comes from our separate study on variability, and neither the 55 per cent nor the 13 per cent can be read without it. We asked the same question at three levels of detail in each of four sectors and recorded every website each answer cited.

Lunasol sells work aimed at how businesses are described in these answers, so the reading here is commercially convenient for its owner, and the notes set out that interest in full.

AI Visibility is owned by Lunasol.

The short version. In the twelve answers Lunasol recorded in Google's AI Mode on 21 September 2026, adding detail to a question changed which websites got cited, but not in proportion to the detail. The identical question, asked again with nothing changed, kept 74 to 77 per cent of its own sources. One qualifier kept 17 of 31, or 55 per cent, below that but not far below. The most detailed version kept 4 of 31, or 13 per cent, which is nowhere near it. With one run at each level we can set these figures beside each other, and we cannot say how much of the movement the added words caused. In two of the four sectors the most detailed question kept none of the plain question's sources at all, and in one sector the question with a qualifier cited the same five websites as the plain one.

What we did

Four sectors, three levels of one question, one run at each level, on one surface, on one day.

Where level 1 came from, stated plainly. The level 1 run of each sector is not new. It is the counted run from our earlier study of which named businesses were also cited, recorded earlier the same day under that study's own protocol, frozen before its first counted run. The same four answers are also run 1 of our separate study on variability. Three articles in this series therefore rest partly on the same four answers, and all three say so.

The four sectors were chosen by a counting rule. The rule, fixed in advance, was every third sector in the order used by our count-only twelve-sector study: business setup consultants, IT support companies, gyms and moving companies.

The three levels, written down before any of them was run. Level 1 is the plain category question, "Which are the best [sector] in Dubai?" Level 2 adds one qualifier. Level 3 adds two or three further details on top of the level 2 qualifier, and the four level 3 questions are not uniform: in the gyms sector one of those details is a district, in the moving sector one is a destination, and in the other two sectors neither appears. Because of that, this article calls level 3 the most detailed version rather than giving it a single label it does not deserve. All eight of the level 2 and level 3 questions were written into the protocol in full before the first of them was asked, so none was tuned after seeing a result. Here they are, verbatim.

The twelve questions Lunasol asked in Google's AI Mode on 21 September 2026, three levels of detail in each of four sectors. All eight level 2 and level 3 questions were written into the protocol in full before the first of them was run.

What was counted. For each run: the total outbound links in the answer that were not Google's own, the distinct websites among them, and how many businesses the answer printed as a bold or heading label at the head of a list entry, which is what named means throughout this article. Then, for each sector, how many of that sector's level 1 websites were cited again at level 2 and at level 3. Two websites count as the same website when they share the same domain name, so the www and non-www forms of one address count once.

The problem with this study, written down before we started rather than found afterwards. In our separate study on variability, the same question asked twice did not return the same sources in most of the sectors we ran. So a difference between two different questions is only evidence of a specificity effect if it is larger than the difference the same question produces by itself. That is why the 74 to 77 per cent figure appears beside every specificity figure in this article, in the tables and in the advice. We wrote that requirement into the protocol before running anything.

What was not done. We did not open a single cited page. We ran each of levels 2 and 3 once, so we have no repeat measurement at either level.

Twelve runs, twelve answers, no failures recorded. No re-run was needed.

How much of the plain question's source list came back at the more detailed levels?

In the four sectors Lunasol ran in Google's AI Mode on 21 September 2026, the plain question cited 31 websites, counted sector by sector. Asking the identical question again kept 24 of those 31, or 77 per cent, and 23, or 74 per cent, on a third asking, which our separate study on variability measured and this study did not. Against that, adding one qualifier kept 17 of the 31, or 55 per cent, and the most detailed version kept 4 of the 31, or 13 per cent.

Lunasol, four sectors in Google's AI Mode, 21 September 2026. The first two rows are from our separate study on variability and are the baseline. The last two are this study. All four are shares of the same 31 websites.

At level 2 the kept share sat below the same-question band, and the move is smaller than the number looks. Fifty-five per cent kept means 14 of the 31 websites dropped out when the qualifier was added. When nothing at all was changed, 7 of the 31 dropped out on the second asking and 8 on the third. So 55 per cent does not mean that one qualifier changed 45 per cent of the answer. Much of that movement is movement this question produces with no change of wording at all, and with one run at level 2 we cannot say how much of it the qualifier caused.

At level 3 almost nothing was kept. Keeping 4 of 31 websites, or 13 per cent, is nowhere near the 74 to 77 per cent the identical question kept by itself. The gap at level 3 is much wider than the gap at level 2, and it is far outside the range the same question produced when we simply asked it again. It still rests on one run per sector, so we cannot say how much of the gap the added detail caused.

Level 2 and level 3 are not two steps of the same size. Level 2 sits close to the same-question figures. Level 3 sits a long way below them. One run at each level cannot tell you the shape of anything in between, and we tested nothing in between.

What did each of the twelve runs cite?

Lunasol, four sectors in Google's AI Mode, 21 September 2026. L1 is the plain category question, L2 adds one qualifier, L3 is the most detailed version. Links are total outbound non-Google links. Websites is how many distinct websites those links point to. Named is how many businesses the answer printed as a bold or heading label at the head of a list entry. Kept is how many of that sector's level 1 websites were cited again at that level. Same question kept is how many came back when the identical question was asked a second and a third time, which our separate study on variability measured and this study did not.

The total website count does not move between level 1 and level 2. It is 31 both times, while 14 of the 31 are not the same websites, against 7 and 8 when the identical question was asked again. That is the finding in one row: across our four sectors the plain question and the question with one qualifier cited the same number of websites and not the same websites.

Did adding a qualifier ever leave the sources unchanged?

In the IT support sector, the question with one qualifier cited all five of the plain question's websites. The two answers Lunasol recorded in Google's AI Mode on 21 September 2026 cited an identical set. It is the only cell in this study where a qualifier left the cited website set unchanged, and it rests on one run at each level.

This is also the cell that shows best why the same-question figures have to sit beside these. In this same sector, on the same day, the identical level 1 question kept five of its five websites when it was asked a second time and two of the five when it was asked a third. Level 2 kept five of five and level 3 kept two of five. In IT support, exactly as much of the plain question's list came back after a wording change as after a plain repeat: five of five and two of five either way. That is a statement about how much of the level 1 list survived, not about the whole of each answer, since the level 3 answer cited thirteen websites against level 1's five. Nothing in our record explains it.

Two sectors went the other way at level 3. In Lunasol's runs, business setup kept none of the ten websites its plain question cited, and gyms kept none of its six. When those same two questions were simply asked again, business setup kept 6 and then 7 of its ten, and gyms kept 5 and then 6 of its six. In those two cells, the most detailed question shared no cited website at all with the plain version of itself.

What else showed up that we were not testing for?

One, the most specific question in the study produced the fewest sources. The most detailed moving question cited four outbound links and four websites. Its plain version cited twenty links and ten websites. Everything else in the study sat between five and twenty-eight links. In our runs, more detail in the question did not mean more sources in the answer. In that cell it meant fewer.

Two, a government source appeared, and it is the only one in any of our runs in English. One of the four websites cited by that most detailed moving question was an official emirate portal. Our count-only twelve-sector study sorted 109 cited websites into categories across twelve plain category questions, counting a website once per question, and recorded no government or official source among them. That study counted the www and non-www forms of one address as two websites, where this one counts them as one, so its count is not folded the way ours is. Our separate study on question language covers the runs that were not in English. We are reporting the appearance, not explaining it, and one website in one run is one website in one run.

Three, the count of businesses named went up at level 2 and then fell. Across the four sectors the answers named 20 businesses at level 1, 24 at level 2 and 14 at level 3. The count did not move in step with the detail in either direction: the most detailed level named the fewest businesses, but the middle level named the most.

What this study cannot tell you

Read this before you quote any number above.

Twelve answers is twelve answers. Four sectors of twelve, one run at each of three levels, one day. None of the shares above is a rate.

Levels 2 and 3 were run once each. We have no repeat measurement at either level, so we cannot subtract the same-question movement from these figures. We can only set them beside it, which is what the tables do. A single level 2 run that happened to land low would look like a specificity effect, and we would have no way to tell the two apart.

The baseline is four sectors on one day too. The 74 to 77 per cent figure comes from our separate study on variability, which ran the same four sectors on the same day. It is the right baseline to use and it is not a large one.

The levels are our wording, not a scale. We chose what counts as one qualifier and what counts as a most detailed version, and our level 3 questions are not uniform: one adds a district, one adds a destination, and two add neither. The eight questions are quoted in full above precisely so a reader can judge them.

One surface, one day, one account, one place. Everything here describes Google's AI Mode on 21 September 2026, signed in to one account, in one browser, on one connection, from one location.

It cannot tell you the cause. We did not measure why the more detailed questions were answered with different sources, and nothing in our record tells us that the added detail is what caused the difference.

It cannot tell you that nothing a business does matters, or that anything does. We tested question wording, not any action a business might take.

Nothing here is a promise. Nothing in this article makes a business named, cited or recommended. Google states that "No one can guarantee a #1 ranking on Google" (Google Search Central). Our article on whether anyone can guarantee a ChatGPT recommendation goes through the position in full.

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 answers recorded, and not legal, financial, medical, regulatory or other professional advice. Nothing here tells 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. None of it tells you what being cited is worth to your business, because this study did not measure that. All five checks below need nothing but a browser and a few minutes.

Check the question your customer would actually type, not just the category. Our own twelve answers all began from the plain category question. In our four sectors, the most detailed version of the question kept 4 of the plain question's 31 websites, or 13 per cent, against 24 and 23 of the same 31 when the identical question was simply asked again. If your customers ask detailed questions, the plain question may be telling you little about them.

Write down the exact wording every time, and keep it. The difference between our level 1 and level 3 is a few words. If you check with slightly different wording each time, you are not comparing like with like, and you will read a wording change as a change in the world.

Check more than one level of detail on your own sector. Run the plain question, then the same question with the qualifier your customers actually use, and compare the two lists of cited websites rather than your impression of the two answers. Write the websites of each answer in two columns, count how many appear in both, and write that count against the length of the first list. That is the whole of what we measured. Then run the plain question a second time, unchanged, and count the same way, so you have your own version of the same-question figure before you read anything into a wording change.

Do not assume the plain question and the detailed question give the same answer about you. In our runs, being cited on one was not always being cited on the other. At the widest gap we tested, in two of our four sectors, the plain question and the detailed question shared no cited website at all, where asking the plain question again kept most of its own list in both of those sectors.

Do not build a plan on twelve answers, 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.

All five are about the wording of the question you ask in Google's AI Mode, which is what this study varied.

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" among the GEO deliverables. GEO is this article's shorthand for work aimed at how a business is described in AI answers.

This study asked questions of Google's AI Mode and recorded which websites each answer cited. That page describes the published audit as "what ChatGPT says about you today", so it is a reading of one business on the assistant named in that line. This study ran on Google's AI Mode instead. Our article on what an AI search audit should cover goes through that in general.

What to do next. Lunasol's published scope for this work is at lunasol.ae/geo-aeo, and our separate article walking through that published service page goes through the scope in detail. If that is the reading you want for your own business, ask for the AI visibility audit that the same page lists.

Common questions about question wording and Google AI Mode

Why is the same-question figure in the headline table? Because without it the specificity figures cannot be read. When the identical question kept 74 to 77 per cent of its own websites, the 55 per cent our four level 2 runs kept between them sits a little below that, and the 13 per cent our four level 3 runs kept sits far below it. Publishing 55 per cent on its own would invite a reader to think that one qualifier had changed almost half the answer. It does not, because the same question repeated changed about a quarter of it.

How many answers is this built on? Twelve. One run at each of three levels in each of four sectors, on one surface, on one day. The four level 1 answers are the same four answers used as run 1 in our separate study on variability, so they are not new here.

Does 13 per cent mean detailed questions are a different search? Thirteen per cent is the share of the plain question's 31 cited websites that survived the most detailed version, against 74 to 77 per cent for the identical question asked again, and we cannot say from four sectors and one run per level that this makes it a different search. What we can say is that at the widest gap we tested, in two of four sectors, the detailed question and the plain question shared no cited website.

Why quote the eight questions in full? Because the levels are our own wording and a reader cannot judge the study without seeing them. If you think our level 2 is really a level 3, the article gives you what you need to say so.

Can I see the raw data? The per-run record exists, with every website in every run. It carries company and domain names, which we do not publish. What this article gives you is the counts taken from that record, which is the per-sector table above. There is no published data file and no separate published report anywhere.

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. Four sectors, chosen by a rule fixed in advance, three levels of one question, one run at each level. The eight level 2 and level 3 questions were written into the protocol in full before the first of them was run. No failures and no re-runs. Counts were read from the rendered page; no cited page was opened. Two websites count as the same website when they share the same domain name.

The shared level 1 run. The level 1 run of each sector comes from our earlier study of which named businesses were also cited, under that study's protocol, frozen before its first counted run. Run 1 of our separate study on variability is the same answer in each case. That is three articles resting partly on four answers, and all three say so. Those three articles are not three independent confirmations of anything.

The baseline was not measured in this study, and it is still ours. The 74 to 77 per cent figure is measured in our own study on variability, on the same four sectors on the same day, and is quoted here rather than measured again. It is not an independent figure and it is not a published standard. If that study is wrong, every comparison in this article is wrong with it.

The website totals are sums, not distinct totals. The 31 at level 1, the 31 at level 2 and the 34 at level 3 are each a sum of the four per-sector counts, and the two 31s are different sets of websites. We did not record whether a website appeared in more than one sector. The named totals of 20, 24 and 14 are sums in the same way.

One method limitation. Businesses were counted as named only when the answer printed a bold or heading label naming them at the head of a list entry. A business named only in running prose was not counted. The rule was applied identically to every run.

Names. No host, domain, company, agency, publication or individual from the study is named anywhere in this article, including the one official portal cited at level 3, which is described by category only. Sector labels are categories, not businesses.

Who publishes this. AI Visibility is owned by Lunasol, whose published work is aimed at how businesses are described in these answers. This protocol, the sector selection rule, the eight added questions and every measure were frozen before the runs they govern, and the level 1 runs were governed by the earlier study's protocol rather than this one, as the method note above sets out. The section where a finding turns into advice is separated and labelled. Our commercial interest runs with a reading in which what an AI answer says about a business matters, because Lunasol's published service page lists an "AI visibility audit, what ChatGPT says about you today" and "Presence in sources AI models trust and quote" among its GEO deliverables. Readers should weigh the numbers above knowing that. The author writes for this publication, which Lunasol owns, and is not independent of either. This article makes no claim about any other provider of these services, 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. OpenAI says of ChatGPT Search that "Search results and citations can be incomplete, outdated, or incorrect" (OpenAI Help Center), 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 and AI Mode are products or publications of Google. All are trademarks of their respective owners.

AI Visibility is owned by Lunasol.