How Much Does a Google AI Mode Answer Change Between Runs? A Lunasol Study

Research conducted by Lunasol. We asked the same question three times in each of four sectors in Google's AI Mode on 21 September 2026. Of the websites the first run cited, 77 per cent were cited again in the second run and 74 per cent in the third, so at most roughly a quarter were not, and 61 per cent were cited again in both. Widen the frame from that first-run list to every website any of the three runs cited, 51 counted within their own sectors, and 19 of the 51, or 37 per cent, were cited in all three runs of their sector, while 21 of the 51, or 41 per cent, appeared in only one. Both figures are true, they have different denominators, and they answer different questions. Lunasol sells work aimed at how businesses are described in these answers, so this finding is commercially convenient for its owner, and the notes set out that interest in full.

AI Visibility is owned by Lunasol.

The short version. Ask Google's AI Mode the same question again and you may not get the same sources. In our four sectors on 21 September 2026, 24 of the 31 websites the first runs cited came back in the second run and 23 came back in the third, which is 77 per cent and 74 per cent. Nineteen of the 31, or 61 per cent, came back in both. Those three figures are all shares of the same 31. Pool every website that any of the three runs of a sector cited, add the four sectors, and there are 51, not 31, because later runs brought in websites the first run never showed. Of those 51, 19 appeared in all three runs and 21 appeared in exactly one. The two percentages are shares of two different pools, 31 and 51, so one is not a stricter version of the other. One sector returned an identical set of five websites twice, then a set of nine that had only two of those five in it. It is a pilot: one surface, Google's AI Mode, one day, four sectors of twelve, three runs each.

Why one run is a sample of one

Each of our studies of what an AI answer cites is a study of one run, unless it says otherwise. If a quarter of the same question's first run sources are gone when it is asked again the same day, then a single run is one draw from something that moves, and any share taken from it carries that movement inside it. That applies to every finding drawn from a single run, including ours.

We use the words draw and pool in this article as a way of reading our own counts, not as a description of anything happening inside Google's AI Mode. We did not measure that and we make no claim about it.

This also gives the two studies that follow a baseline. Our study of what happens when a question gets more specific, and our study of the same question asked in Arabic, both measure their effect against the first-run figure in this article, the 74 to 77 per cent of a first run's cited websites that came back in a later run of the same question. Without it, a change between two different questions cannot be told apart from the change you get by asking the same question twice.

What we did

Four sectors, one question shape, three runs each, on one surface, on one day.

Where the three runs came from, stated plainly. Run 1 of each sector is the counted run from our earlier study of which named businesses were also cited, which asked this same question across twelve sectors earlier the same day, under that study's own protocol, frozen before its first counted run. That is not the count-only twelve-sector study compared at the end of this article, and the two counted websites by different rules. Runs 2 and 3 were added here, under this study's protocol, frozen before those two runs. Each protocol was frozen before the runs it governs. Not all three runs of a sector were run under one protocol, and we would rather say so than let a reader assume otherwise.

The four sectors were chosen by arithmetic, not by how they behaved. The rule, fixed in advance, was every third sector in the twelve-sector order used by our count-only twelve-sector study of which categories of source get cited: business setup consultants, IT support companies, gyms, moving companies. The rule is arithmetic and was applied to the frozen sector order, not to how any sector had behaved.

One question shape. Every run used the same sentence with only the sector term changed: "Which are the best [sector] in Dubai?" Runs 2 and 3 of a sector were consecutive in one browser session, on the same connection, signed in to the same account. Run 1 was recorded earlier the same day, in the earlier study's pass across twelve sectors. We did not record the elapsed time between any of them, so no interval is stated anywhere in this article.

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 gave an emphasised entry label to. Then, for each sector, three set operations fixed before the data was seen: how many websites were cited in all three runs, how many in exactly one, and the size of the union across the three. Two websites count as the same website when they share the same domain name, the part of the address that identifies the site, so the www and non-www forms of one address count once and not twice. No judgement is involved in those three counts. They are set arithmetic on the recorded lists.

Two counting cautions, because they affect the figures above. First, 31 and 51 are sums of the four per-sector counts. We did not record whether a website appeared in more than one sector, so neither figure is a count of distinct websites across the study, and we do not call them that. Second, our count-only twelve-sector study counted the www and non-www forms of a domain as two websites where this study folds them, which matters for the comparison at the end and nowhere else.

What was not done. We did not open a single cited page. So this is a record of what was cited, not of what those pages say. We also did not record the order in which an answer listed anything, so nothing here is about ranking.

Twelve runs, twelve answers, no failures recorded, so no re-runs were needed. In the later of the two twelve-sector sets compared at the end, the extraction script was tested once on the interior fit-out prompt before that study's protocol was frozen. That test was discarded and is not in that dataset, and the interior fit-out row in that table is the counted run made under the frozen protocol.

How much came back? The two honest answers

Between 74 and 77 per cent of the first run's cited websites came back in a later run of the same question, and 37 per cent of the 51 websites the three runs cited between them were cited in all three. There are two ways to count this and they give different numbers, because they have different denominators. Both are below, with their denominators, because a share without its denominator is not a finding.

Counting forward from the first run. In the four sectors Lunasol ran in Google's AI Mode on 21 September 2026, the first runs cited 31 websites, counted within their own sectors. The second runs cited 24 of those 31 again, and the third runs cited 23 of them. That is 77 per cent of the first run's websites back in the second run, and 74 per cent of them back in the third. Nineteen of the 31, or 61 per cent, were back in both. This is the set of numbers to use when the question is "if I check today, how much of what I see will still be there when I check again?"

Counting across all three runs. Pool every website any of the three runs of a sector cited, add the four sectors, and there are 51. Nineteen of those 51 websites, or 37 per cent, were cited in all three runs of their sector. Eleven of the 51 were cited in exactly two. Twenty-one of the 51, or 41 per cent, were cited in exactly one run and never again. These are the numbers to use when the question is "how many different websites turned up across all three runs, and how much of that total was in every one?"

The gap between 74 per cent and 37 per cent is not a contradiction, and the two are not shares of the same thing. The first number starts from a fixed list, the 31 websites the first runs cited, and asks how much of it came back. The second starts from all 51 and asks how much of that was there every time. Business setup is the plain case. Its first run showed 10 websites, but across the three runs 12 different websites appeared in that sector, so 2 of the 12 were never in the first run at all. Those 2 can only lower the second number, because the first number counts nothing the first run did not show. The other 2 of that sector's 4 cited-once websites were in the first run and did not come back, and those lower both numbers.

Lunasol, twelve Google AI Mode answers, three runs in each of four sectors, 21 September 2026. The first three columns hold one figure per run and are not summed across sectors. The last four count websites across the three runs of that sector, and the last column is the sum of the four per-sector totals rather than a count of distinct websites across the study. Cited in all three, in exactly two and in exactly one add to it: 19 plus 11 plus 21 is 51.

The sector that stood still, then jumped

In Lunasol's study, the first two IT support runs in Google's AI Mode cited an identical set of five websites, and the third cited a set of nine with only two websites in common with them. The first two runs also produced ten outbound links each; the third produced thirty-one.

Seven of that sector's twelve different websites were cited in one run and never again, and those seven are a third of the study's 21 cited-once websites. It is also the sector where the count-only study recorded fourteen websites, which are twelve organisations once the www and non-www forms of the same address are folded as this study folds them, against five in each of the first two runs here and nine in the third.

We did not measure why, nothing in our record tells us, and a reason with nothing behind it would be worth nothing to you. What we can say is that in this sector, on this day, two runs of the same question cited an identical set of five websites and a third cited nine that shared only two of them.

Of the four sectors, the moving companies question had the most websites cited in all three runs, seven, though that sector's union was also the largest at sixteen. Measured as a share of its own union, three of the four sectors land close together: the gyms sector at five of eleven, moving at seven of sixteen and business setup at five of twelve. Those three are too close to rank on twelve answers. IT support was the outlier in the other direction at two of twelve.

No sector returned the same set of websites three times.

What else changed between runs that we did not count

The protocol told us to record what we saw, not only what we were counting.

One, in the third moving run the answer stopped naming businesses altogether. Its emphasised labels were move-size categories rather than company names, so under the rule fixed in advance no business in that run counted as named, against five in each of the two runs before it. The answer still cited thirteen websites. That is a change in the shape of the answer, and it is all we recorded about it.

Two, a price band appeared in two of the three moving runs and not the third. We are not reproducing the figures. A figure with nothing neutral behind it becomes an anchor the moment a reader sees it.

Three, a business recorded as an arguable case in the first of these three runs gained an entry label. In that run one business in the gyms sector was discussed in the answer's prose without a label of its own, which made it an arguable case we reported rather than resolved. In the later runs of the same question, the same business carried an entry label. The rule was not changed. We mention it because it shows that what counts as named can itself move between runs.

The wider check, on twelve sectors

Read the limits before the table. Our count-only twelve-sector study ran the same twelve prompts on the same day, ahead of these runs, but it recorded how many websites each answer cited and not which ones, so no overlap between that study and this one can be computed. It also counted the www and non-www forms of a domain as two websites where this study folds them together. So the table below compares counts only, across two different counting rules, and it can show that the numbers moved without showing what moved. In the four sectors this study covers, the later column is this study's run 1, so those four rows are not additional answers.

Lunasol, two sets of twelve Google AI Mode answers, one prompt per sector in each set, both on 21 September 2026. Counts only, and the website columns are not like for like; the limits are stated in full above the table.

Across the two sets in the table, the link count differs in eleven of the twelve sectors. Nine sectors differ on the website count. The two sets counted websites by different rules, but that difference changes only the IT support row, where the earlier fourteen is twelve once the forms are folded. Ten differ on the count of businesses named. In the earlier set, two of those counts, car service centres and moving companies, were derived by matching the answer's entry pattern rather than read entry by entry, so those two rows carry a method difference as well as a count difference. Two more of that set's named figures, accounting firms and nurseries, are composite readings rather than plain counts of named entries, so the accounting row, which is one of the two that match, matches across two different readings. The 105 is our own sum of that set's twelve figures and not a total that study publishes. Three sectors returned the same website count both times, which is not the same as returning the same websites, and this study cannot tell you whether they did.

What this study cannot tell you

Read this before you quote any number above.

Twelve answers is twelve answers. Four sectors of twelve, three runs each, one day. None of the shares above is a rate and none should be treated as stable.

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 says nothing about any other assistant, any other day, or what another person would see.

It cannot tell you the cause. A run that draws on three times as many links as the one before it may be doing something different or may be doing the same thing differently. Nothing on the page says which, and we did not measure it.

Three runs cannot give you a distribution. With three observations per sector we can say the set moved. We cannot say how often it moves, by how much when it does, or whether it settles.

Every comparison here is measured against run 1. The 77 per cent, the 74 per cent and the 61 per cent all count runs 2 and 3 against run 1, and run 1 was recorded under the earlier study's protocol rather than this one. The two protocols used the same measures and the same website counting rule. They did not use the same setting: run 1 was taken in a pass across twelve sectors, while runs 2 and 3 were taken as a dedicated pair. Runs 2 and 3 are the only pair recorded under one protocol, and the table above fixes their overlap: 21 of the 30 websites run 2 cited across the four sectors, or 70 per cent, were cited again in run 3. Like the 31 and the 51, those are sums of per-sector counts. That is one pair in four sectors, so we do not put it beside the headline figures, but it is the closest thing here to a same-setting comparison and it sits a little below the run 1 figures rather than above them. Which way the confound cuts is worth saying. An uncontrolled difference in setting is more likely to add to the movement we measured than to take it away, so the quarter that did not come back should be read as a ceiling on how much the same question moves by itself rather than as an estimate of it. The same applies to the 51 and to the 37 per cent taken from it. A larger figure for movement is the reading most convenient for the owner of this publication, which is why this sits here and not in a footnote.

Runs taken the same day are not the same as runs a week apart. Runs 2 and 3 were consecutive in one session. Run 1 was recorded earlier the same day, in the earlier study's pass across twelve sectors. We did not record the gap between any of them. Whether the same question drifts more or less over longer gaps is a different study and we have not run it.

It says nothing about ranking. We recorded which websites were cited, not the order of anything.

It cannot tell you that nothing a business does matters. We measured how much the cited set moved between identical runs. We tested no action a business might take, so nothing here shows that any such action works, and nothing here shows that it does not. Both readings would be going past the data.

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.

Never check once. A single run of a question about your market is one draw. Run the question three times, close together, and write down, for each run, which websites the answer linked to. Compare the three lists, not your impression of the three answers. In our four sectors, between 74 and 77 per cent of the 31 websites our first runs cited came back in a later run. We cannot tell you what share would come back in your sector on your day, and one check of your own cannot tell you either.

Treat an absence from one run as a question, not an answer. Twenty-one of the 51 websites in our four sectors appeared in exactly one run of three. If your site is missing from a check, that is consistent with being in the pool and not drawn that time. It is also consistent with not being in the pool. One run cannot tell you which.

Treat a presence in one run the same way. The logic runs in both directions and this is the direction people forget. Being cited once is not evidence that you are reliably cited, and a single check of any kind is one draw.

Write down the date, which account you were signed in to, and the city you were in, every time. We held all three the same and the answers still moved. We did not test what happens when they differ, so if any of the three change between your own checks, write that down before you read anything into the gap.

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 need nothing but a browser and a few minutes.

How this study relates to Lunasol's published process

Lunasol's published process ends with a step called Track, described on our service page (Lunasol) as "Numbers, every quarter". The same page lists "AI visibility audit, what ChatGPT says about you today" among the GEO deliverables. GEO is this article's shorthand for work aimed at how a business is described in AI answers.

Repeated measurement is the part of that process this study bears on directly. A reading taken once is a reading taken once, whoever takes it. The published page lists Track as the last of its four steps, after Audit, Signals and Content. The published audit is described as "what ChatGPT says about you today", which is a reading of one business on the assistant named in that line, while this study ran on Google's AI Mode. 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 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, and for Track, which the same page lists as the last of its four process steps.

Common questions about how much Google AI Mode answers change

Which number should I quote, 74 per cent or 37 per cent? Whichever matches your question, and say which. 74 per cent is 23 of the 31 websites the first runs cited, the share still there in the third run, in Lunasol's four sectors in Google's AI Mode on 21 September 2026. 37 per cent is 19 of the 51 websites that appeared at all, the share that was in all three runs.

Did Google AI Mode return the same set of websites every time? In no sector did all three runs return the same set. One sector did return an identical set of five websites in two of its three runs. In our four sectors on 21 September 2026, between 74 and 77 per cent of the first run's cited websites came back in a later run of the same question.

Does this mean AI answers are random? We are not going to use that word in either direction. Three runs per sector cannot establish it and cannot rule it out, which is the same reason we say elsewhere that three runs cannot give you a distribution. What we can say is that a clear majority of the first run's sources came back in each of the two later runs taken singly, and one sector returned an identical set twice.

How many answers was this? Twelve. Three runs in each of four sectors, on one surface, on one day. The wider comparison at the end sets an earlier set of twelve answers from the same day beside a later set of twelve, one prompt per sector in each, and it compares counts only. In the four sectors this study covers, that later set is this study's first runs, so four of its twelve answers are already counted in the twelve.

Why only three runs? Because three is enough to show that the set moves and not enough to say how it moves. A study that could say how much and how often is a larger piece of work, and we would rather publish the small thing accurately than the big thing loosely.

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 separate 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 runs each, one prompt shape. Run 1 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, and not from the count-only twelve-sector study compared at the end. Runs 2 and 3 were added under this study's protocol, frozen before those two runs. 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 set arithmetic was checked per sector and in total: cited in all three, in exactly two and in exactly one sum to the union in every sector, and 19 plus 11 plus 21 is 51.

Two counts that are sums, not distinct totals. The 31 and the 51 are sums of the four per-sector counts. We did not record whether a website appeared in more than one sector, so neither is a count of distinct websites across the study.

The wider comparison is counts only. The count-only twelve-sector study recorded how many websites each answer cited and not which ones, so no overlap between the two sets can be computed. It also counted the www and non-www forms of a domain as two websites, where this study folds them. Both facts are stated beside the table.

One method limitation. Businesses were counted as named only when an emphasised entry label named them. A business named only in running prose was not counted. The rule was applied identically to every run.

The related studies. Our study of what changes when a question gets more specific, and our study of the same question asked in Arabic, both use the four sectors and the first runs reported here as their baseline.

Names. No host, domain, company, agency, publication or individual from the study is named anywhere in this article. 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 and every measure were frozen before the runs they govern, and run 1 of each sector was 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 this article's central finding rather than against it. The finding is that one reading is not enough to conclude anything, and Lunasol sells a process whose published last step is repeated measurement, listed on the service page as Track, "Numbers, every quarter". 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.