A UAE business starting ninety days of AI search work should spend the first week measuring rather than changing anything: write down ten to twenty questions your customers would actually type, ask them, and ask one of them twice in a row so you know how much the answer moves by itself. Spend the rest of the first month on the two things that are either true or not true about your website, which are whether AI crawlers can reach it and whether it states the facts a buyer asks for. The second month goes on the pages themselves, service by service, and on the second language if your customers use one. Spend the third month measuring again with exactly the same questions, logging what arrived as enquiries, and writing the first report. That is the whole plan, and the order is the part that matters, because every step after week one is unreadable without week one. Two sections below say what you will have on day 90 and what you still will not know, and name the three things that make the ninety days impossible to read. This article sequences the work and points at the article in this series that sets out each piece in full, rather than repeating any of them.
This is general guidance on planning your own marketing, not legal, financial, tax, accounting 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, and nobody can promise that it will.
The short version: the work in a ninety-day plan is not controversial and you can find a list of it anywhere, including from the AI answers themselves. What almost nobody puts in the plan is the measurement that makes the ninety days readable. 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, with every website total there a sum of per-sector counts rather than a count of distinct websites across the study. So an answer that looks different on day 90 may be different for reasons that have nothing to do with anything you did. Google itself says that after a change, "some changes can take effect in a few days, while others could take several months" (Google Search Central), and ninety days is at the short end of that. Week one is where you find out how much your own questions move, and that number is what day 90 gets read against.
Lunasol sells the AI visibility work this plan describes, and the notes set out that interest in full.
AI Visibility is owned by Lunasol.
Why is week one a measurement and not a change?
A ninety-day plan is a before-and-after, and a before-and-after needs a before that you can trust. 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. Across the three runs, 51 websites appeared in total, of which 19, or 37 per cent, appeared in all three, and 21, or 41 per cent, appeared in exactly one. Our study of the identical question asked three times reports those figures, the sectors and the counting rules in full.
Read the second of those figures again before you plan anything. Twenty-one of those 51 websites, two in five, turned up once and then never again across the three askings of an unchanged question that Lunasol ran in four Dubai sectors in Google's AI Mode on 21 September 2026. If you take one reading on day 1 and one on day 90, a large part of what you see moving is the reading itself moving. The fix is not a better tool. The fix is to ask one question twice in a row in week one, count both answers the same way, and write down what share of the first answer's websites the second answer carried. That single figure is your noise floor, and every comparison you make in the next twelve weeks is read against it.
Google says the same thing in its own terms about its own product. Its guidance on investigating a traffic drop says that "Small fluctuations in position can happen at any time (including moving back up in position, without you needing to do anything)", and that "In general, you'll likely want to wait a few weeks to analyze your site in Search Console again" (Google Search Central). A plan that checks weekly and reacts weekly will spend ninety days chasing movement that was going to happen anyway.
What did Google AI Mode itself propose for the first 90 days, in two runs on 21 September 2026?
Lunasol asked Google's AI Mode what a business should do in its first 90 days to appear in AI search answers, twice in one session on 21 September 2026, under a protocol frozen before the first run. The first answer set out 11 steps as emphasised entries, each one a bold label or a heading of its own, and cited 11 websites from 27 outbound links. The second set out 12 steps and cited 10 websites from 24 links. Of the 11 websites the first answer cited, 9 were cited again in the second, 9 of 11, and the second added one the first had not.
Not one of the 23 step labels was identical across Lunasol's two runs of the same question in Google's AI Mode on 21 September 2026. Eleven labels in the first answer, twelve in the second, and under the counting rule we froze before running, which matches labels as written, zero appeared in both. The closest pair was a step called "Increase Fact and Data Density:" in the first answer and "Increase Fact Density:" in the second.
Read instead by what each step asks you to do, the two plans are close to the same plan. That grouping is a reading we made after the runs and not a measure we set in advance. Every one of the first answer's eleven steps has a counterpart in the second, and the second has one more only because it splits a single step about comparison tables and FAQ markup into two. Both put crawler access first and a baseline audit second. Both end with a re-audit. Both divide the ninety days into three blocks of thirty. That these two answers agree on the shape of the work is a fact about two answers on one day and nothing more.
Both answers told the reader to take a baseline before changing anything, which is what the rest of this article argues for too. The first called its second step "Conduct an AI Search Audit:" and its last "Re-Audit and Refine:", and told the reader to re-run the day 1 audit at the end. The second called its second step "Run a Manual AI Baseline Audit:" and told the reader at the end to re-run the initial audit across several assistants. The plan below does the same thing, for the reason set out in the section above rather than because two answers said so.
Neither of the two Google AI Mode answers Lunasol recorded on 21 September 2026 told the reader that the thing being audited moves between runs by itself. Not in the first, not in the second, not in any step, not in any sentence. So a reader following either plan would take one reading on day 1 and one on day 90 and read the difference as the result of ninety days of work, with nothing to tell them how much of it the question produces on its own.
One of the two answers promised a quantified outcome inside the ninety days, and named no source for it. That answer said, of businesses that adopt what it called a systematic GEO strategy, and attributing the figures to unnamed "industry data", that they would "see 2x as many citations" and 3 to 9 times higher conversion rates "within 90 days". The wording is reported here as a record of what that answer contained and for no other purpose. The rendered answer gave no study, no publisher and no date for that industry data, and the range in it, written in the answer with a dash this publication does not use, is given here as 3 to 9. The first answer made no outcome claim at all. We cannot check the claim, we have measured nothing that supports it, and no study in this series has connected a citation to a sale.
Both runs carried Google's own line, verbatim, "AI responses may include mistakes."
Days 1 to 30: measure, then fix what is either true or not
Week 1. Freeze the questions and take the baseline. Write ten to twenty questions in the words a customer would type, covering the services you actually sell, in the languages your customers actually use. Print them, date them, and do not change them for the rest of the ninety days. Ask each one once on the surface your customers use, meaning the one AI product you are going to measure on, and record, for each answer, the date, the language you asked in and whether you were signed in, then how many outbound links it carried, how many distinct websites those links pointed to and which websites they were, whether your business was named, and whether your own website was among the links. Week 9 repeats all of those conditions, so a reading you cannot describe is a reading you cannot repeat. Our article on how to check what AI says about your company sets out the run design in full, and its free test sheet is where the answers go.
Week 1, the part almost nobody does. Ask one question twice in a row. Same wording, same session, straight away. Count the second answer the same way, and write down how many of the first answer's websites the second answer carried, as a count over a count. That is your noise floor, and it costs you fifteen minutes.
Week 2. Find out whether AI crawlers can reach the website at all. This is a yes or no about your own server and it does not depend on anybody's opinion. When we looked for the robots.txt file of each of thirty-five UAE accounting and tax firm websites in September 2026, twenty of them cited in AI answers and fifteen not, 28 of the 34 files we could read named no AI crawler at all, in either direction. Four of the 20 cited websites named any AI crawler, and 2 of the 14 reachable uncited ones did. Almost nobody had made a decision, which means the decision is still available to you. Our article on whether AI search can access your website is the checklist, and it also covers the part most people get wrong, which is that one crawler decides search and a different one decides training.
Week 3. Put the facts a buyer asks for on the website, as text. Not in an image, not in a PDF, not only in a form. When we checked twenty UAE accounting and tax company websites that AI answers had cited, in September 2026, all 20 named their services and a city or emirate, but only 9 of the 20 stated any price, 12 of the 20 a licence or regulator status, and 7 of the 20 their opening hours. Those are the gaps to close first, because they are cheap and they are factual. Publish a price only where you can stand behind it, and state a licence or regulator status only where you hold it and only in the wording your own licence uses. Our article on whether your website explains who you are lists the eighteen facts and how to write each one, and it comes with a free checklist.
Week 4. Write down what you changed and when. One row per change: the date, the page, what changed, and who did it. This is the most boring hour of the ninety days and it is the only record that will let anybody, later, say which month had a change in it. Without it, day 90 produces a difference and no candidate explanation.
What day 30 looks like. You have a frozen question set, one reading of each question, one repeat, a decision about crawlers, the basic facts published as text, and a log of what you changed. Nothing in that list is a result. All of it is the apparatus that lets day 90 mean something.
Days 31 to 60: the pages, in the order your money is in
Weeks 5 and 6. Rewrite your service pages, starting with the one that sells most. One page per service, written to the six blocks rather than as a brochure paragraph, because a page built that way contains the facts an answer would have to lift and a brochure paragraph does not. Our article on how to write service pages sets out the six blocks in order, with a worked rewrite, and it is the longest piece of practical work in the ninety days.
Week 7, only if you sell products. Fix the product information. Price, availability, delivery, returns and variants are the fields that go out of date fastest and the fields a shopping answer needs most. The twenty-eight fields are in our article on UAE e-commerce product information, with a free checklist to work from. If you do not sell products, spend week 7 on the second-biggest service page instead, and leave every other week where it is, because the gap between the two readings is the only thing making them two readings.
Week 8. Decide about the second language, and if the answer is yes, treat it as a separate measurement. Do not translate your question set and assume the reading transfers. When Lunasol asked the same category question in Arabic in four Dubai sectors in Google's AI Mode on 21 September 2026, the Arabic runs cited 50 websites against the English 31, counting sector by sector, and only 15 of those English 31 came back in Arabic, 48 per cent, against the 24 and 23 of those same 31 that the identical English question kept when it was simply asked again, 77 and 74 per cent. A separate list means a separate reading, its own questions, and its own repeat. Our article on English and Arabic AI search covers how to build a matched pair, and our study of what changed in Arabic reports the figures in full. Use the Arabic wording a real customer would type, and if you do not read Arabic, ask an Arabic-speaking customer, colleague or supplier what they would type and have them read the answers with you.
What day 60 looks like. Your pages say what you do in words that can be lifted, your facts are text, your crawler decision is made, and you have not measured anything since week one. That is deliberate. The next reading is the one that counts, and taking extra readings in between will mostly show you the movement you already measured in week one.
Days 61 to 90: read it, log it, write it down
Week 9. Run the questions again, unchanged, and run the repeat again. Same words, same surface, same signed-in state, same language, same counting rules. Then ask the same question you repeated in week 1 twice in a row again, because your noise floor is a property of the question and the surface on the day, not a constant you measured once in week one.
Week 10. Ask every enquiry how they found you, and write the answer down. Ask openly, keep the note against the enquiry record you already hold rather than starting a new file about people, and take your own professional advice on what you may record and for how long. Not the attributed click. The answer from the person. This is the only thing that connects an AI answer to a human being who contacted you, and no analytics product will do it for you. Our article on measuring AI referral to enquiry sets out the five stages, says which of them you can count reliably, and comes with a free template for the log.
Week 11. Write the first report. Our article on what a monthly AI visibility report should contain sets out the nine sections it needs and publishes a free workbook with a tab for each one, so this week is filling in what weeks 1 to 10 already produced rather than deciding anything.
Week 12. Decide what the next ninety days is, from the report rather than from a feeling. Count how many of the websites your week 1 answers cited came back in your week 9 answers, as a count over a count, and put that share next to the noise floor you wrote down in those same two weeks. If the week 1 to week 9 share sits inside your noise floor, the honest reading is that nothing measurable has happened yet, and the decision is whether to keep going, not whether the work failed. If something moved further than your noise floor, the change log you started in week 4, which covers every change from week 2 onwards, is the list of candidate explanations, and none of them is proof.
One platform number belongs in that first report. Google's own documentation describes a generative AI performance report in Search Console, and what that documentation describes it as publishing is impressions, for AI Overviews and AI Mode counted together, grouped by pages, countries, dates and devices. That documentation names no clicks, no click-through rate, no position and no queries. Google states that "As of August 31, 2026, we've rolled out these insights to all websites worldwide" (Search Console Help). If your website is already set up there, this is the only number in the whole ninety days that comes from the platform rather than from your own counting.
What will you have on day 90, and what will you still not know?
You will know whether your website can be reached, and whether it states the facts. Those are properties of your own server and your own pages. They were either true on day 1 or they were not, and they are either true now or they are not. This is the part of the ninety days that is fully under your control and fully verifiable.
You will know what two readings of a frozen question set looked like, with a noise floor beside each one. That is a real measurement with a date on it. It is not a rate, it is not a trend, and two readings is two readings.
You will not know whether the work caused anything. Nothing in this plan is an experiment with a control. Google's own guidance on investigating a change says that after a change to a site, "some changes can take effect in a few days, while others could take several months" (Google Search Central), so a page rewritten in week 6 may not have been read by anything that matters by week 12.
You will not know what a citation was worth. No study in this series has connected being cited to money. If you want to know what arrived, the enquiry log from week 10 is the only instrument in the plan that touches a human being, and ninety days of it is a small number of rows.
Google's own suggested question for anybody selling this work is worth asking yourself. Among the questions Google recommends putting to an SEO before hiring one is "What kind of results do you expect to see, and in what timeframe?" (Google Search Central). A plan that cannot answer it in counts and dates is not a plan.
What three mistakes make these ninety days unreadable?
One, changing the questions. It feels harmless and it destroys the comparison. When we added a single qualifier to the same category question in four Dubai sectors in Google's AI Mode on 21 September 2026, the more detailed question kept 17 of the 31 websites the plain question had cited, 55 per cent, and the most detailed version kept 4 of 31, 13 per cent. Our study of what a more detailed question changed reports those runs in full. Change a question and you have started a new measurement, whatever the calendar says.
Two, one run per question. A single reading carries the movement you measured in week one and shows you none of it. Two runs of one question, twice in the ninety days, is the minimum that makes the other readings interpretable.
Three, no record of what changed and when. Without the change log, day 90 gives you a difference and no list of candidate causes, and the temptation is then to credit whatever you happen to remember doing.
What this plan cannot tell you
It is a sequence, not a method, and not a standard. Every piece of work in it is set out in full in a sibling article, and nothing has been validated against anything. The claim here is about order, not about efficacy.
The measured figures in it come from small studies on one surface on one day. The retention band of 23 and 24 of 31 websites, 74 and 77 per cent, the 17 of 31, 55 per cent, and 4 of 31, 13 per cent, for question detail and the 15 of 31, 48 per cent, for Arabic are all four Dubai sectors, in Google's AI Mode, on 21 September 2026, quoted here rather than measured again. Every website total among them is a sum of per-sector counts rather than a count of distinct websites across a study. Your own noise floor, measured in week one on your own questions, is the only one that applies to you.
The two website checks quoted in weeks 2 and 3 are smaller still. Thirty-five websites and twenty websites, in one professional category, in one country, in one month, read from a browser on one connection. They illustrate where the gaps usually are. They are not rates for UAE businesses and no figure in them is a finding about any company.
The live check in this article is two answers. One question, one surface, one day, two runs. It says what two answers contained and nothing about what such answers usually contain.
No plan 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), and 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). Ninety days of careful work does not put anybody in an answer, and nothing in this article says otherwise.
Documentation moves. Every description of a Google or OpenAI product above repeats that product's own published documentation as fetched on 21 September 2026. Check the current version of any page linked here before relying on it.
Where to start this week
Write the questions down today. Ten to twenty, in the words a customer would type, in the languages your customers use. Date the list. That is the single step everything else in the ninety days depends on, and it needs nothing but a text file.
Ask one of them twice in a row before you change anything. Count both answers, write the two counts down as one number over another, and you have a noise floor for your own questions on the day you measured it.
Open your robots.txt and look at it. It takes two minutes and the answer is a yes or a no.
Put your opening hours, your price or price basis, and any licence or regulator status you actually hold, on the website as text. Three facts, one afternoon, and they are the ones most often missing. Publish only what your own records support, use the wording your own licence or registration uses, and take your own professional advice before describing a regulated status.
Then start the log. A dated row for every change from here on, so that in twelve weeks the report has something to work with.
How Lunasol runs the first ninety days
Lunasol's published process starts where this plan starts. The service page (Lunasol) calls the first step "The AI census" and describes the audit as "AI visibility audit, what ChatGPT says about you today", which is the day 1 reading in this plan under a different name. Among the GEO deliverables that page lists is "Presence in sources AI models trust and quote", which is the unit this plan counts in weeks 1 and 9: the websites an answer cites, counted as websites.
The last step on that page is named track and is described as "Numbers, every quarter" (Lunasol). A ninety-day plan and a quarterly reading are the same rhythm, and the week 11 report is what a quarter is made of.
If you would rather have the first ninety days run for you than run it yourself, the audit on that page is where that starts. Our separate article on that GEO and AEO service page goes through it in detail.
Common questions about a 90-day AI search plan
Is ninety days enough to see a result? It is enough to produce two readings with a noise floor and a change log, which is a result in the sense that it is a fact about your business with a date on it. Whether the readings move is a different question, and Google's own guidance allows for changes taking several months.
What if I have no website changes to make? Then weeks 2 to 7 are short, and the time goes into week 1 before the clock starts: more questions, and the repeat run on more than one of them. Once week 1 is frozen, extra readings before week 9 will mostly show you the movement you already measured. A bigger question set measured twice is worth more than a smaller one measured once.
Can I compress this into thirty days? You can do the work in thirty days. You cannot compress the two readings, because the gap between them is the only thing making them two readings.
Which surface should I measure on? The one your customers use. If you do not know, ask the next five enquiries. That answer is worth more than any guess about market share.
Do I need a tool for this? No. A browser, a text file and a spreadsheet are enough for the whole ninety days. Tools help when the question set is large or when several people are counting, and neither is true in the first quarter.
What if my sector is not in your studies? Then the figures quoted here tell you what the surface did in four Dubai sectors on one day, and your own week one repeat tells you what it does in yours. The second one is the one to plan against.
What should I do on day 91? Repeat it, with the same questions. The second ninety days is where the first one becomes readable.
Method, sources and notes
The live check. One question, "What should a business do in its first 90 days to appear in AI search answers?", 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 and no cited page was opened. Steps were counted only where the label appeared as an emphasised entry in the rendered answer, the same rule used in the other studies in this series; steps mentioned only in running prose, table row labels and follow-up format offers were not counted. The grouping of steps by what they ask the reader to do, 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 outcome claim in the second run, which this publication cannot check and does not endorse. The range inside that claim is written in the rendered answer with a dash this publication does not use, so the claim is quoted in part and the range is reported outside the quotation as 3 to 9. No website cited in either answer is named in this article, and no step or claim recorded in either answer is attributed to anyone.
The quoted study figures. The retention figures of 24 of 31 websites, 77 per cent, and 23 of 31, 74 per cent; the 51 websites across three runs with 19 in all three, 37 per cent, and 21 in exactly one, 41 per cent; the 17 of 31, 55 per cent, and 4 of 31, 13 per cent, for question detail; the 15 of 31, 48 per cent, and the 50 websites against 31 for Arabic are all measured in Lunasol's own studies published in this series and are quoted here rather than measured again. Each is four Dubai sectors, in Google's AI Mode, on 21 September 2026. The four-sector studies share a common first run and are not independent confirmations of each other, and each of those articles says so. Every website total among them is a sum of per-sector counts rather than a count of distinct websites across a study.
The two website checks. The crawler-access figures in week 2 and the company-information figures in week 3 come from two earlier checks published in this series, on thirty-five and twenty UAE accounting and tax firm websites respectively, in September 2026, read from a browser on one ordinary connection with one attempt per file. Both are single-sector illustrations of a method rather than studies, both say so in full in their own articles, and neither names any company.
The product documentation. Every description of Google Search Console, the generative AI performance report, Google's AI features, Google's guidance on investigating a traffic drop and Google's guidance on hiring an SEO, and of ChatGPT Search, repeats that product's own published documentation as fetched on 21 September 2026. The Lunasol service page quoted above was read on the same date. Every quotation in this article is from the version of that page published on that day, and 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 this work is worth ninety days of a business's attention, 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 plan above knowing that, and knowing that a reader who follows this plan is doing work Lunasol is paid to do for other businesses. The plan is written to be run by a business on its own, with free templates, and the limits section says plainly what ninety days of it cannot establish. 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 two plans 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 is general information current at September 2026. It is not legal, financial, tax, accounting or professional advice of any kind, and nothing in it 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 and AI Overviews are products or publications of Google. All are trademarks of their respective owners.
AI Visibility is owned by Lunasol.
