A monthly AI visibility report should contain the question set it used, printed in full; the runs it did, with the date, the surface, meaning which AI product was asked, and the conditions on each one; what those runs cited; a repeat of at least one question inside the same session, so the month has its own noise floor, meaning its own measure of how far the answer moves when nothing has been changed; how many businesses the answers named and how many of those also had their own website cited; the platform numbers that actually exist, which as of September 2026 means impressions in Google Search Console and not clicks; what you changed on your own website; what arrived as enquiries; and a plain statement of what the report does not claim. Nine sections. The order matters, because the fourth one, the repeat, decides whether anybody can read the third, the count of what the answer cited. Both of the report designs we recorded from Google's AI Mode led with a single percentage, and neither attached a number to it. A single percentage is the one thing a month of this work cannot honestly produce. A free workbook with a tab for each of these nine sections is published with this article.
This is general guidance on reporting your own marketing, not legal, financial, tax or professional advice, and nothing here describes what any UAE rule requires. Nothing in it makes a business shown, named, cited or recommended in any AI product.
The short version: the source list behind an AI answer moves on its own. In our own runs on Google's AI Mode in four Dubai sectors on 21 September 2026, the identical question asked a second time kept 24 of the 31 websites it had cited the first time, 77 per cent, and a third asking kept 23 of the same 31, 74 per cent. That is one day, one surface, four sectors, and every website total there is a sum of per-sector counts rather than a count of distinct websites across the study. It means a report that shows you went from eleven cited websites to nine has not yet shown you anything, because the answer produces a change of that size by itself. The fix is not a better score. The fix is to run one question twice in the same session every month, publish both, and let every other number be read against that pair. A report built that way is duller and worth keeping.
Lunasol sells the AI visibility auditing and tracking that this article argues is worth doing and worth writing down every month, so the design here is commercially convenient for us, and the notes set out that interest in full.
AI Visibility is owned by Lunasol.
Why a monthly AI visibility report needs a noise floor
A month-on-month change can only be read against how much the answer moves when nothing changes at all. Lunasol ran the identical question three times in one session in each of four Dubai sectors in Google's AI Mode on 21 September 2026. Counting sector by sector, the four first runs cited 31 websites between them. The second asking of the same question cited 24 of those 31 again, 77 per cent. The third cited 23 of the same 31, 74 per cent. Run 2 against run 3 overlapped by 21 of the 30 websites run 2 cited, 70 per cent. Every one of those totals is a sum of four per-sector counts rather than a count of distinct websites across the study.
That movement sets a floor under every monthly comparison. If seven of 31 websites can drop out of a list inside one session on one asking, and eight of the same 31 on another, then a report comparing August with September is reading a difference that includes a month of real change, that much ordinary movement, and no way to tell which is which. Publishing the second number without the first is not a small omission. It is the difference between a measurement and a decoration.
Table 1. How much a cited source list moved in our own runs. Lunasol, four Dubai sectors in Google's AI Mode, 21 September 2026, counting sector by sector against the 31 websites the first runs cited.
| What changed between the two runs | Websites kept from the first 31 | Share |
|---|---|---|
| Nothing. The identical question, asked again | 24 of 31 | 77 per cent |
| Nothing. The identical question, asked a third time | 23 of 31 | 74 per cent |
| One qualifier added to the question | 17 of 31 | 55 per cent |
| The most detailed version of the question | 4 of 31 | 13 per cent |
| The same question asked in Arabic | 15 of 31 | 48 per cent |
Lunasol, four Dubai sectors in Google's AI Mode, 21 September 2026. The first two rows are the noise floor. The last three are what changing the question, or the language, did against it.
The table is also a warning about the question set. Changing one word of a question is not a free edit. In those runs, adding a single qualifier left 17 of the 31 websites in place, 55 per cent, and the most detailed version of the question left 4 of 31, 13 per cent. A report whose question set drifts between months is not comparing months. It is comparing questions.
What did Google AI Mode itself propose, in two runs on 21 September 2026?
Lunasol asked Google's AI Mode what a monthly AI visibility report for a business should include, twice in one session on 21 September 2026, under a protocol frozen before the first run. The first answer set out 13 report sections as emphasised entries, each one a bold label or a heading of its own, and cited 8 websites from 23 outbound links. The second answer set out 15 sections and cited 12 websites from 32 links. All 8 of the websites the first answer cited were cited again in the second, 8 of 8, and the second added 4 the first had not cited.
One section label appeared, word for word, in both of the two Google AI Mode answers recorded on 21 September 2026. That was the headline score. Under the counting rule we froze before running, which matches labels as written, 13 labels in the first answer and 15 in the second produced exactly one string that both runs used. Read instead by what each item measures, which is a reading we did after the fact and not a measure we set in advance, most of the first answer's items do come back in the second under different names. Those are the share-of-voice item, the evidence log, the competitor item, the source mix, the technical readiness check and the action list. Four items appeared only in the second answer, among them a section on misinformation risk and a gap summary. Two appeared only in the first.
Both of the two Google AI Mode answers Lunasol recorded on 21 September 2026 led with a percentage, and neither attached a number to it. The first described the headline score as "the overall percentage of commercially relevant, tracked prompts where the brand appeared", and the second as "The overall percentage of tracked commercial prompts where your brand appears". Neither said how many prompts, how many runs of each prompt, or over what period. That is the figure this article argues should never be published on its own, and it was the first line of both proposed reports.
Neither of the two Google AI Mode answers recorded on 21 September 2026 proposed a section in which the same question is asked more than once. The second answer came closer than the first: it described AI engines as "non-deterministic" and recommended a report focused on "four-week trend directions rather than daily fluctuations". The first raised none of it. So on one of the two runs, a reader following the answer would have built a monthly report with a headline percentage, no denominator and no repeat. We are reporting two answers on one day. Two answers is two answers.
The second answer also argued against the shape of the report it had just proposed, and it is worth quoting. It said, in the rendered answer, that "A useful report must never rely on a standalone proprietary score; every metric must connect to visible reality and an explicit owner." We agree with it, and it is the one sentence across both runs that would have changed the design of the report the same answer proposed.
What should a monthly AI visibility report contain? The nine sections, in order
One. The question set, printed in full, with the date it was frozen. Not a description of the questions. The questions, in the words they were typed, with the language of each one marked. Ten to twenty is a workable set for a single business. The set is fixed for at least a quarter, and when a question changes, the report says so on the same page and stops comparing that question with its earlier self. Our article on how to check what AI says about your company covers how to build the set, and our article on what an AI visibility audit should include covers the one-off audit that this monthly report is not.
Two. The run log. One row per run: question, surface, date, time, whether the account was signed in, the location, the language, and anything unusual. This is the section people skip, and it is the section that decides whether anyone can repeat the month. Two reports with the same score and different signed-in states are not comparable, and only the run log would ever show it.
Three. What each answer cited. For each question: how many outbound links the answer carried, how many distinct websites those links pointed to, and how many of last month's websites for that same question came back this month. Three numbers, each with its denominator. Websites, not links, are the stable unit, because one website can be linked four times in one answer. An outbound link here means a link in the answer that goes to a website other than the one you asked on, and two links count as the same website when they share the same domain name, so the www and non-www forms of one address count once.
Four. The repeat, run in the same session. At least one question from the set, asked twice in a row on the day of the run, with both answers counted the same way as every other run. By the same session we mean asking the question again straight away, in the same browser, with the same signed-in state, rather than on another day. Our own record does not say whether each repeat was typed into the same conversation or into a fresh one, so decide which of those you will do, do it the same way every month, and write down which you chose. This is your own noise floor for the month, and it goes in the report as a figure, not as a caveat. Without it, section three is unreadable. With it, a reader can see at once whether this month's movement is bigger than the movement your own question produces in a single session.
Five. Named and cited, counted separately. Being named in an answer and having your own website cited in that answer are two different events, and a report that merges them will overstate one of them. In twelve Dubai sectors on Google's AI Mode on 21 September 2026, Lunasol recorded 78 businesses named across the twelve answers, of which 64, or 82 per cent, also had their own website cited in the same answer. Thirteen of those 78, 17 per cent, were named without their own website being cited. One further pair was arguable and was held out of both counts.
Six. The platform numbers that exist. As of September 2026 that means the generative AI performance report in Google Search Console, and it means impressions. The next section sets out precisely what the generative AI performance report covers and what it does not, because a report that promises platform-verified clicks from AI answers is promising something the platform does not publish.
Seven. What you changed on your own website this month, with dates. Pages published, pages rewritten, facts corrected, schema added, anything that changed what a crawler would find. Without this column the report cannot even suggest a cause, and with it you at least know which months had an input. Our article on how to write service pages and our article on whether your website explains who you are cover what usually goes in this column.
Eight. Enquiries, and what people said when you asked how they found you. The count of enquiries, by route, and what each one answered when asked. Not attributed clicks. Answers from people. Our article on measuring AI referral to enquiry sets out the five stages and the template that holds them, and this section of the monthly report is the short version of that log.
Nine. What this report does not claim. Four or five lines, in the report, every month: that these are counts and not rates; that the sample is what it is; that no figure here proves a cause; that being cited has not been shown to be worth a stated amount to this business; and that nobody can promise a position in an AI answer. This is the section that makes the other eight defensible, and it is the first one that gets cut.
How to write the change column so the number can be read
Every share carries its denominator in the same line. Not "citation share up 12 per cent". 9 of 14 this month, against 7 of 13 last month. The denominators are different, and printing them is how the reader finds that out rather than being told a story about a percentage.
Compare like with like or do not compare. Same question, same wording, same surface, same language, same number of runs, same signed-in state. If any of those changed, the row says so and the comparison is dropped for that question this month. Our own runs are the reason for the strictness: one added qualifier left 17 of the 31 websites in place, 55 per cent, and the same question asked in Arabic left 15 of 31, 48 per cent. Either of those would read as a dramatic month if the question had quietly changed.
Say when a change is inside your own repeat range. Take two invented figures. If this month's question kept 8 of last month's 12 websites, and the repeat you ran inside one session kept 8 of 12 as well, then the honest line is that the month-on-month figure sits inside the range the question produced in a single session. That line is what lets a reader tell the months that moved from the months that did not.
Report three numbers, not one. What came back, what is new, what dropped. A source list that kept 8 of 12 and added 5 new websites is a different month from one that kept 8 of 12 and added none, and a single retention figure hides that completely.
Never pool unlike questions into one figure and call it typical. Twelve questions across four service lines do not have a meaningful middle. Report them as a table and let the reader see the spread. Where a pooled figure genuinely helps, say what it is a sum of, in the same sentence.
Write the raw counts even where the share is the interesting part. Shares of small numbers move violently. Two websites out of four is 50 per cent, and so is fifty out of a hundred, and only one of those tells you anything.
The one figure most AI visibility reports get wrong
The figure is the single visibility score, usually written as a percentage of prompts where the brand appeared. Prompt is the word those reports use for what this article calls a question. Both of the answers we recorded put it first. It is attractive because it fits in a subject line, and it fails for four reasons at once, each of which is fixable and none of which is fixed by making the score more precise.
It has no denominator on the page. A percentage of a prompt set is meaningless unless the set is printed, and the sets that produce flattering scores are the ones nobody prints.
It is built from one run of each prompt. Our own repeats moved 7 of 31 websites on one asking and 8 of the same 31 on another, inside a single session, and a score built on one run of each prompt carries that movement with no way to see it.
It merges being named with being cited. In the twelve Dubai answers Lunasol counted, 13 of the 78 businesses named, 17 per cent, had no link to their own website in the same answer. A score that counts both as one event is counting two different things.
It cannot be checked by the person receiving it. If the report does not print the questions, the runs and the counts, the score cannot be recomputed by anyone, including the person who made it. A number that cannot be recomputed is not a measurement, whatever decimal place it is given to.
The replacement is not a better score. It is the four numbers underneath it. How many questions, how many runs each, how many of them named you, and how many cited your website. Those four can be checked, and a reader who wants a headline can make one from them.
What Google's own report gives you, and what it does not
Google Search Console publishes, as of 21 September 2026, a generative AI performance report, and what it publishes is impressions. The report covers, in Google's words, "AI Overviews, AI Mode" on Google Search, and the two are counted together rather than separately (Search Console Help). Google defines an impression there as how many times links to your site were shown to a user in a generative AI feature on Google Search. The page names impressions and no other metric: no clicks, no click-through rate, no position, and no search queries.
It reached everybody recently. Google states that "As of August 31, 2026, we've rolled out these insights to all websites worldwide" (Search Console Help). If your September report is the first one with this section in it, that is why.
What the Help page says you can group it by. Pages, countries, dates and devices (Search Console Help). For a UAE business that is more useful than it sounds, because the country grouping is the only place in a monthly report where the platform itself tells you where the impressions were.
Two limits belong in the report, not in a footnote. Google says "The newest data can be preliminary, meaning it's still being collected and might change in the next few hours", so a month closed on the first of the month can move after you send it. And Google says it "doesn't include data from experiments in Search Labs" (Search Console Help), so anything a reader saw in an experimental surface is not in this number.
Google's documentation describes no way to separate clicks from AI features, and the report should say so. Google states that "Sites appearing in AI features (such as AI Overviews and AI Mode) are included in the overall search traffic in Search Console" and reports them in the Performance report inside the Web search type (Google Search Central). So the overall Search clicks in your report already contain whatever AI features sent you, mixed in with everything else, and Google's documentation describes no filter that pulls them apart. A monthly report can carry an impressions line from the generative AI report and a total Search clicks line from the performance report. It cannot carry an AI clicks line. Any report that does is producing that number somewhere other than Google.
There is a separate report for Discover. Keep it in its own row if you use it. We checked what Google publishes and nothing else, so if your customers use another assistant, what that one publishes about your website is a question this article does not answer. Merging Search and Discover into one AI number is the same mistake as merging named with cited.
A worked month, using figures we measured
Here is one question filled in, with real recorded numbers. Two of the nine rows are marked not applicable rather than invented, and the table says why in each of them. The runs are our own, on the business setup consultants question, arranged the way a monthly report would arrange them.
Table 2. One question written up as a monthly report entry. Lunasol, Google's AI Mode, 21 September 2026, three runs of the identical question in one session.
| Report section | What this month's row says |
|---|---|
| 1. Question | The plain category question for business setup consultants in Dubai, in English, frozen 21 September 2026 |
| 2. Run log | Google AI Mode, 21 September 2026, signed in, one browser, one connection, one location, three runs in one session |
| 3. What the answer cited | Run 1: 21 outbound links, 10 distinct websites |
| 4. The repeat, same session | Run 2: 14 links, 7 websites, 6 of run 1's 10 came back, 60 per cent. Run 3: 12 links, 8 websites, 7 of run 1's 10 came back, 70 per cent |
| 5. Named and cited | Businesses named as emphasised entries: 6 in run 1, 6 in run 2, 5 in run 3. Named and cited counted separately, never merged |
| 6. Platform numbers | Not applicable. We own no website in these answers |
| 7. Website changes this month | None. Nothing was changed on any website by us |
| 8. Enquiries | Not applicable. This was a study, not a client month |
| 9. What this does not claim | Three runs of one question on one day. No rate. No cause. No value per citation |
Read row 4 before row 3, and the rest follows. The repeat in the same session kept 6 of the first run's 10 websites, 60 per cent, and the third run kept 7 of the same 10, 70 per cent. That is this question's floor. It means that next month, a month-on-month figure near the 6 of 10 or the 7 of 10 this question produced in a single session tells you nothing at all, and a figure of 2 of 10 would be worth an hour of somebody's time. Without row 4 there is no way to know which of those you are looking at, and the report would have treated both as news.
The free monthly report template published with this article
The workbook comes with this article as a free download, as 28-monthly-ai-visibility-report-template.xlsx. It opens in Excel, Numbers or Google Sheets. In Google Sheets, open it through File, Import and choose Replace spreadsheet so the formulas and the formatting come through. Read the Start here tab, then fill the Question set tab once. Everything else is filled a row at a time as you run.
Monthly report template developed by Lunasol Lunasol, which owns AI Visibility, built this workbook for the article and released it free, to use and to pass on. Ten tabs: a Start here tab and one for each of the nine sections of the report. The repeat tab, which is section four, works out what share of the first run's websites came back in the second. Fill the yellow cells only. On tabs one to eight the grey row under the header is an example, so leave it alone and start below it.
The first twenty minutes with it. Open the Question set tab and write your questions, one per row, in the words you would type them, with the language marked. Freeze them for the quarter by putting today's date in the frozen column. Then open the Run log and record the run you are about to do, before you do it. Run the first question, count the links and the distinct websites, and put those in the Citations tab. Then ask the same question again, immediately, and record that as a second run. The Repeat tab will tell you what share of the first run's websites came back, and that figure is the one to read everything else against for the rest of the month.
What it deliberately does not compute. There is no score cell. There is no cell that turns nine counts into one number, because the article you are reading is an argument that the one number is the problem.
What this article and this report cannot tell you
A reporting design is not a measurement standard. Nothing here has been validated against anything. It is a way of writing down what you did that lets somebody else check it, and that is the whole claim.
The noise floor figures are four sectors on one day. They come from our own study of how much a Google AI Mode answer changes between runs, which our article on that study reports in full, on one surface, on one day, in four sectors of twelve, and they are not a published rate for anything. They are quoted here rather than measured again. Your own repeat, run inside your own session, is the only floor that applies to your questions, and that is why section four asks you to measure it yourself rather than borrow ours.
The live check in this article is two answers. One question, one surface, one day, two runs. It says what two answers contained. It says nothing about what such answers usually contain, and a third run could have produced a different list again, which is the same point the rest of the article makes.
None of this measures what a citation is worth. No study in this series has connected a citation to money. A report can tell you that you were cited in six of ten answers with a stated denominator. It cannot tell you what that was worth, and anybody who fills that gap for you is filling it with an assumption.
No report changes how the surfaces work. Google states that "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary" (Google Search Central). Google also states that "No one can guarantee a #1 ranking on Google" (Google Search Central). OpenAI says of ChatGPT Search that "Search results and citations can be incomplete, outdated, or incorrect" (OpenAI Help Center). A monthly report is a record of what happened, not an instrument that makes anything happen.
Documentation moves. Every description of a Google Search Console report above repeats that product's own published documentation as fetched on 21 September 2026. It changes without notice. Check the current version of any page linked here before you rely on it.
Where to start with next month's report
Start with the repeat, not with the score. Pick one question you already ask. Ask it twice in a row, in one session, on the same day. Count the distinct websites in both answers and write down how many of the first answer's websites came back in the second. That single figure, with its denominator, is what every other number in the month is read against.
Print the questions. Whatever else is in the report, if the questions are not in it in the words they were typed, nobody can check anything and next month's comparison is a guess.
Add the Search Console impressions line. If your own website is already set up in Google Search Console, this is the only platform-published number in the whole report, and it puts a source outside your own counting into the document. If it is not set up there, that is the step before this one.
Put the four raw numbers where the score used to be. How many questions, how many runs each, how many named you, how many cited your website. If somebody wants a headline they can build one, and it will be a headline anybody can recompute.
Then do it again next month with the same questions. Two months of a report built this way is where the first real reading starts, and the question set you froze this month is what makes the second month worth reading. Write next month's run log entry before you run anything, the way you did this month.
How Lunasol reports this
Lunasol's published process begins with an audit. The service page (Lunasol) calls it "The AI census" and describes the audit as "AI visibility audit, what ChatGPT says about you today". Among the GEO deliverables that page lists is "Presence in sources AI models trust and quote", which is the same unit as section three of the report above: the websites an answer cites, counted as websites.
The reporting rhythm on that page is quarterly, and its last step is named track and described as "Numbers, every quarter" (Lunasol). A monthly report of the kind set out here and a quarterly reading sit together well: the monthly rows are the record, and the quarter is where three months of rows are long enough to read as a direction.
Our separate article on that published service page goes through it in detail. If you would rather have the reading done for you than build the workbook yourself, the audit on that page is where that starts.
Common questions about monthly AI visibility reporting
How many questions should the set contain? Ten to twenty for a single business, covering the services you actually sell, in the languages your customers actually use. Fewer than ten and one answer swings everything. More than twenty and the repeat runs stop happening, which costs you more than the extra questions bought.
How often should the questions be re-run? Monthly is a reasonable rhythm for the report, with at least one question run twice inside the same session every month. Our article on how to check what AI says about your company covers the run design itself.
Can I compare my report with somebody else's? No. Different question sets, different surfaces, different signed-in states and different counting rules produce numbers that look comparable and are not. Compare your months with your own earlier months, and only where the question was unchanged.
What if the report shows no movement at all? Write that down and keep it. A month where the source list held still is a real result, and it is also the month that makes the next move readable.
Should the report include competitors? You can count how many businesses an answer named and how many of them had their own website cited, and that is a count. Naming individual competitors in a document that circulates is a decision with consequences beyond measurement, and it is not one this article can make for you.
Is a single visibility score ever acceptable? As a summary of four printed numbers on the same page, yes. As the only number, no, for the four reasons set out above.
What is the smallest useful version of this? One question, two runs on the same day, the websites counted, the questions printed, and one line saying what it does not claim. That fits on half a page, and every figure in it can be recomputed by the person who receives it.
Method, sources and notes
The live check. One question, "What should a monthly AI visibility report for a business include?", asked twice in Google's AI Mode on 21 September 2026, signed in, one browser, one connection, one location, the second run immediately after the first. The protocol, the question, the measures and the tie-break rules were written and frozen before the first counted run. No failures and no re-runs. Counts were read from the rendered page. No cited page was opened. Report sections were counted only where the label appeared as an emphasised entry in the rendered answer, which is the same rule used in the other studies in this series; labels mentioned only in running prose were not counted. The grouping of items by what they measure, rather than by their label, is a reading made after the runs and is marked as such in the text. Both runs are reported, including the sentence in the second run that argues against the report design the same answer proposed. No website cited in either answer is named in this article, and no report design recorded in either answer is attributed to anyone.
The quoted study figures. The retention figures of 24 of 31 websites, 77 per cent, 23 of 31, 74 per cent, and 21 of 30, 70 per cent, the 17 of 31, 55 per cent, the 4 of 31, 13 per cent, and the 15 of 31, 48 per cent, the per-sector business setup figures in Table 2, and the 78 businesses named with 64 cited and 13 not, are all measured in Lunasol's own five studies published in this series and are quoted here rather than measured again. Each of those studies is four sectors or twelve sectors, on one surface, on one day. They are not independent of each other: the four-sector studies share a common first run, and every one of them says so. If those studies are wrong, the arguments in this article that rest on them are wrong with them. Every website total quoted is a sum of per-sector counts rather than a count of distinct websites across a study, and that wording travels with the figure wherever it is used.
The product documentation. Every description of Google Search Console, the generative AI performance report, the Performance report and Google's AI features repeats that product's own published documentation as fetched on 21 September 2026, from the Search Console Help page on the generative AI performance report and from the Google Search Central page on AI features and your website. The Google Search Central page on whether a site needs SEO, the OpenAI Help Center page on ChatGPT Search and the Lunasol service page quoted above were fetched on the same date, and every quotation in this article is from the version of that page published on 21 September 2026. Those pages change without notice.
Who publishes this. AI Visibility is owned by Lunasol, whose published work is aimed at how businesses are described in AI answers. Our commercial interest runs with a reading in which what an AI answer says about a business matters enough to be measured and reported every month, because Lunasol's published service page (lunasol.ae/geo-aeo) lists an "AI visibility audit, what ChatGPT says about you today" and "Presence in sources AI models trust and quote" among its GEO deliverables, and lists a tracking step described as "Numbers, every quarter". Readers should weigh the design above knowing that. The workbook published with this article computes no score and recommends against publishing one, which is a position that makes our own reporting harder to flatter rather than easier. 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, and the report designs described in the live check section are the content of two AI answers rather than the work of any identified party.
General information. This article and the workbook published with it are general information current at September 2026. They are not legal, financial, tax, accounting or professional advice of any kind, and nothing in them describes what any UAE authority requires or permits. Nothing here makes a business shown, named, cited or recommended in any AI product, and nobody can promise that it will.
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, Google Search Console, AI Mode, AI Overviews and Google Discover are products or publications of Google. Excel is a product of Microsoft, Numbers of Apple, and Google Sheets of Google. All are trademarks of their respective owners.
AI Visibility is owned by Lunasol.
