English and Arabic AI Search: What UAE Brands Should Test

If you sell to both English-speaking and Arabic-speaking customers in the UAE, you have two audiences asking machines about your category, and you are testing one of them. The fix is not to translate your English prompt set. Arabic is a separate test: its own question pairs matched on intent rather than on wording, your brand name written out in both scripts, and the same thirteen fields recorded for every run in both languages.

In September 2026 we put one matched pair to Google's AI Mode: one question in each language, on the same day, in the same tool, on one connection, about the same Dubai sector. It was a method demonstration rather than a study. In that single pair, eight businesses were named in the English answer and six in the Arabic one, four of them in both. That is what a recording sheet captures, and nothing about any market.

The short version: build pairs, not translations. Record both wordings, because no pair is ever exact. Add your brand's Arabic forms to the test: in the one Arabic answer we logged, the assistant transliterated every business name itself, and never asked which spelling the company preferred. Record the same thirteen things for every run, in both languages, and freeze that recording sheet before the first run. Two runs cannot tell you which language serves you better, and neither can this article.

AI Visibility is owned by Lunasol.

Does asking in Arabic get a different AI answer? Why Arabic is a separate test

In the one pair we ran, the two answers differed on almost everything we recorded: the businesses named, the shape of the answer, the kind of page cited and the number of links out. They matched on two things: both carried a map card, and both ended with three questions back, two of which asked the same thing. That is one pair and not a pattern, and it is the reason to measure the two languages separately rather than assume one stands in for the other. Between them the two answers named ten businesses and agreed on four. One built a comparison table, the other an advice list. And the visible source cards were not the same kind of thing: on the English side, entries labelled with a platform name; on the Arabic side, three cards with Arabic headlines and Arabic summaries.

There is a reason to expect this on the page side too. Google states plainly how it reads language: "Google uses the visible content of your page to determine its language. We don't use any code-level language information such as lang attributes, or the URL" (Google Search Central). A page that is English with an Arabic menu is an English page. The Arabic customer's question is being answered from somewhere. In the one Arabic run we logged, every visible source card carried an Arabic headline and an Arabic summary.

Arabic is in scope for the features people actually see. Google's help pages list Arabic among the languages in which AI Overviews are available, and the United Arab Emirates among the countries and territories (Google Search Help). That is a statement about availability, and about a different feature from the AI Mode we ran; it does not say that answers differ by language, and we are not going to put that claim in its mouth. What we can say is that in our own single pair, they differed.

How to build a matched English and Arabic question pair

A matched pair is two questions that a buyer in each language would actually type, aimed at the same moment in the same decision. It is not one question translated twice.

The set starts from the questions rather than from the page. Lunasol, which owns this publication, lists question research among its published answer engine deliverables, described there as "Question research, what your customers actually ask" (Lunasol). We fixed our own pair before anything was run for the same reason: a set built after you have seen an answer is not a set.

Here is the pair we froze before running anything:

The matched pair we froze before running anything, one Dubai sector, September 2026

Those two are not translations of each other, and we recorded that in the protocol before we ran them. The English asks which one to use. The Arabic asks which is best. They are not the same question, and the reason is specific: English "should" has no single Arabic equivalent, and you do not "use" a service centre in Arabic the way you do in English, so a word-for-word rendering comes out stilted. An idiomatic Arabic version can get close using "choose" or "recommend" instead, but each of those is a different string and therefore a different test. We picked one wording and wrote down why. So you match on intent, you keep both wordings, and you say so in the report.

Four shapes cover most of a buying decision, and each needs its own pair:

Four question shapes, and the English and Arabic wording each one needs

The Arabic gaps are not interchangeable, and this is the part that gets missed. The comparison row needs an indefinite phrase in the gap. The price row wants the definite version instead. The category row takes either. City names carry whatever article they naturally have, so one city takes none and another takes one, and neither should be added or stripped by rule.

Which means you should not fill those gaps yourself unless you write Arabic. Ask for finished sentences instead.

How to get your Arabic test questions written if nobody in your company writes Arabic

If nobody in your company writes marketing Arabic, do not commission a translation: brief an Arabic-speaking colleague on the buying moment, ask for finished sentences, then have a second Arabic speaker tell you in English what the sentence says. Here is that in five steps, and a sixth for the reader who has nobody at all.

Ask a colleague, not a translator. The person you want is an Arabic speaker who talks to your customers: a salesperson, a receptionist, a support agent, somebody at a distributor. Hand a translator your English sentence and no context, and you get that sentence back in Arabic, which is the exact failure this section is about. Brief the intent first, and whoever writes it will write the right question.

Brief them with the intent, not the English. Tell them the moment in the decision: someone has decided they need this and is choosing who to call. Do not show them the English question first, because they will render it.

Ask for finished sentences, not fills. For each of the four shapes, ask for one complete Arabic question with your own sector word already inside it. Ask for each sentence whole rather than for one phrase you reuse across them. The comparison and price shapes need your sector word in different forms, and which form goes where is not something you can judge without reading Arabic.

Ask two or three people separately. Where their wordings converge you have the string customers use. Where they diverge you have found a variant worth testing, which is a result rather than a problem.

Check it without reading Arabic. Give the finished string to a second Arabic speaker and ask what it says in English. If the round trip does not match the intent you briefed, it is the wrong string.

If you have nobody at all, use the four rows above with your own city, your own sector word and your own brand name written into the gaps. Machine translation is acceptable for the city and the sector word, which are single nouns. Leave your brand name in Latin script for this run rather than machine-translating it, because a machine will often translate a name that should be transliterated, and the section on name variants below is where that gets settled properly. It is the sentence around them that must not be machine-translated. None of the rows works with a gap left in it, and the brand row in particular turns into a question about the words "brand name". Then write on the sheet that the pair was template-built rather than customer-written, and that no Arabic speaker checked the form of the sector word or the Arabic form of your name. That is a weaker test and an honest one, and it is much better than testing nothing in Arabic.

What to record for every run: the thirteen fields, in both languages

Freeze the sheet before the first run, not after it. This is the part we got wrong ourselves: we finished the English run, moved on, and only then realised we wanted the list of cited domains in both languages so we could count the overlap. We had not recorded it, and recovering it would have meant a second English run, which is a different run and not a comparison. So the number was never captured and we do not report it.

One word first, because the article uses several and you only have to record one. Cited page means a page the answer links out to, counted once per address. That is the thing you copy into the sheet. Everything else in this article, the source cards, the hosts and the address paths, is a way of looking at the same set.

The thirteen fields to record for every run, in both languages

How to match the names without reading Arabic. Copy the bracketed Latin form when the answer gives you one; ours gave one for every business. When it does not, paste the Arabic name into its own column and match on the string rather than on the meaning, because identical strings match whether or not you can read them. Log the ones you could not resolve instead of dropping them, because a dropped name inflates the overlap and flatters the result.

Cited pages: copy the address, never retype it. Four of the seven distinct address paths in our Arabic answer carried Arabic characters in the address itself. Retyping one is how you get a row that matches nothing.

Set the interface language once and hold it. For your first pair, run both halves with the same browser and account interface language, so the only thing that changes between them is the prompt. It is not what an Arabic-speaking customer's own session looks like, and that is the point: you are isolating one variable. Later, add a run of the Arabic prompt on an Arabic interface as an extra row rather than as a replacement, and you will be able to see which of the two changes moved the answer. Record the choice either way.

Five traps that break an English and Arabic question pair

Five things turn a matched pair into two unrelated questions: a literal translation, the wrong Arabic register, place names in one script only, conditions you did not record, and right-to-left text that changes between your clipboard and the transcript.

1. The literal translation. Covered above, and it is the one that catches everybody. Translate the intent, not the words, and record both wordings.

2. Register. The prompts above are Modern Standard Arabic. That is the safer default for a written set, because the Gulf dialects have no standardised spelling, so the same spoken question can be typed several ways and you lose the ability to repeat a run exactly. It is not how anyone talks at home, and it is not the only thing people type: dialect wording gets typed too, and some customers type Arabic in Latin letters. Treat each as a separate string and a separate test, because one register does not stand in for the others. If most of your customers are Emirati rather than Arabic-speaking residents from elsewhere, dialect wording will be a bigger share of what you are missing.

3. Place and area names. In the Arabic answer we logged, districts arrived in both scripts: the Arabic name with the Latin form in brackets after it. Your test should ask in both forms, because your customer might type either, and your page should state both for the same reason.

4. The conditions you test from. Location and interface language are variables, and if you do not record them you cannot compare anything. OpenAI says of ChatGPT search that it "may use an approximate location based on your IP address to provide relevant local results" (OpenAI Help Center). We ran both halves on an English browser interface, which is itself a condition worth recording. The recording sheet above says which way to set it and why. How to set up and record a run properly in general is covered in our article on how to check what AI says about your company, and it is worth doing it that way rather than improvising.

5. Right-to-left text that changes on the way in. This one nearly cost us a run, and it would have if we had trusted the screen. Our own entry method turned the Arabic prompt into a string with every space removed: twenty-seven characters where the frozen prompt has thirty-three. That is a fact about our tooling rather than a defect in any product, and we did not send it. We re-entered the prompt a different way and checked it against the frozen version before letting it run.

You can run the same check without reading a word of Arabic, and it takes three steps. Paste your frozen prompt in, never retype it. After sending, copy the prompt back out of the transcript. Compare the two, either by character count or by pasting both into any tool that shows differences. If they differ by so much as one character, throw the run away and start again. Thirty-three characters becoming twenty-seven is visible to anybody.

Arabic brand name transliteration: the name variants worth listing in both scripts

If you have not published an Arabic form of your brand name, an assistant may write one for you. In the Arabic answer we logged, every business was named in Arabic with its Latin name in brackets immediately after it. The assistant transliterated the names itself. It did not wait for anybody's permission, and it did not ask which spelling the company preferred.

That means one of two things is true of your brand. Either you have published an Arabic form of your name and that is the one in circulation, or somebody else's transliteration is. There is usually more than one plausible transliteration, because Arabic spelling offers more than one defensible way to write the same name. You are not expected to produce these yourself. This is the list you hand to whoever writes your Arabic, so that they generate the variants instead of you guessing at them:

• p and v are written several ways. A p is normally written ب and a v is normally written ف. There are also borrowed letters kept for foreign names, پ for p and ڤ for v, and a v is sometimes written و. So a brand with either sound is already in circulation in more than one spelling before anybody decides anything. Put them all on the sheet.
• The definite article attaches. Your name may appear with ال in front of it or without, and the two are different strings.
• Hamza and alif. Alif with hamza above (أ) and alif with hamza below (إ) are routinely typed as a bare alif (ا). They are not interchangeable in formal spelling, but in typed Arabic the hamza is dropped often enough that a name beginning with one of them is already in circulation in more than one form before anybody thinks they have made a mistake. The same goes for final ya (ي) and alif maqsura (ى), which are swapped at least as often.
• Taa marbuta. A name ending in taa marbuta (ة) is often typed as haa (ه). It is partly habit and partly the keyboard: on most Arabic phone layouts ة is a long press away from ه.
• The legal suffix. ذ.م.م, short for ذات مسؤولية محدودة, sits roughly where LLC sits in an English name. You will see it typed without the stops as ذ م م, and you will also see the longer ش.ذ.م.م. These are spelling variants you may come across, not a guide to which designation applies to your company: copy whatever appears on your own trade licence rather than choosing between them. A name is often searched without any of them.

So list them. This is the sheet, and it is short enough to fill in one sitting:

The brand name variant sheet: the name forms to list in both scripts

Then publish the forms you want used, in text, on the page, meaning the trading and brand forms you have chosen. Not the misspellings: those belong on the test sheet rather than on the site. And not a rendering of your registered legal name that differs from the one on your licence. A form you have published as text on your own site is at least a form an answer can take from you. A form that exists nowhere on your site can only be taken from somebody else, and in the answer we logged the assistant wrote one itself.

What one English and Arabic pair returned in Google's AI Mode

This was a method demonstration, not a study. One question pair, one tool, one day, one run each. It shows what the sheet above records. It cannot show which language is better served, and the English against Arabic study is separate work that we have not run. What follows is what the two runs recorded, and nothing beyond it.

Two Google AI Mode answers to one matched pair, one Dubai sector, September 2026

Four things are worth pulling out of that.

Ten businesses, four in both. The two halves of one pair, same tool, same day, agreed on fewer than half the names. A brand that is in one and not the other would look fine on any report that only tested one language.

The two answers cited different kinds of page. The English run's visible cards were labelled with the platform name rather than a publisher, and read like business-profile entries. The Arabic run's were an Arabic classifieds portal, an Arabic motoring title and a question-and-answer site, all dated between 2021 and 2025.

The answers wanted different things from the customer. Both ended with three questions. Two matched: what car, and what service. The third did not. The English run offered a choice between two named industrial areas. The Arabic run asked which area of Dubai the customer preferred. Verbatim, it asked "ما هي المنطقة المفضلة لك في دبي لتوجيهك لأقرب ورشة؟"

The second tool produced nothing, and we are telling you. We had planned four answers and got two. The other assistant we tried returned an opening fragment and citation chips on three attempts and would not send on a fourth, in both languages alike. We do not name it and we attribute no cause. The rule this produces is worth more than the incident: record the runs that fail, and do not top up the tool that is working to make the numbers match.

Bilingual website basics: separate URLs, one language per page, hreflang both ways

A bilingual test tells you where the gaps are. Closing them is page work, and Google's own guidance on multilingual sites is specific about how to do it.

• Give each language its own URL. Google recommends different URLs for each language version rather than adjusting the content by cookie or browser setting, and it says of the alternative: "Don't use IP analysis to adapt your content. IP location analysis is difficult and generally not reliable" (Google Search Central).
• Keep one language per page. "You can help Google determine the language correctly by using a single language for content and navigation on each page, and by avoiding side-by-side translations."
• Translate the content, not just the furniture. Google warns that "Translating only the boilerplate text of your pages while keeping the bulk of your content in a single language ... can create a bad user experience." An Arabic menu over English body copy is the version of this we come across most often, and by Google's own description it is the version that undercuts the page.
• Link the versions to each other, both ways. Google's instruction on hreflang is to "tell Google about the variations of your content, so that we can understand that these pages are localized variations of the same content", and it is strict about the mechanics: "Each language version must list itself as well as all other language versions", and "If two pages don't both point to each other, the tags will be ignored" (Google Search Central).
• Use codes that exist. "Only language codes listed in ISO 639-1 and region codes listed in ISO 3166-1 Alpha 2 are supported".
• State the facts in both languages, not just the marketing. The company facts an answer quotes are the ones written as text, and our article on the company information your website should include lists them. Each one needs an Arabic version if you want the Arabic answer to have something to quote, with the same caution as in English: a person's details go up only with that person's agreement, and licence and registration details stay as your own licence states them. The same goes for the service page itself, which our article on writing a service page breaks into six blocks.

That is the page-side half of the job. The test tells you where to point it.

The control set to ask any supplier for in a two-language test

The control set below is what to ask any supplier for before commissioning a two-language test: named people on both language sides, an intent-matched prompt set with both wordings written down, the same number of runs per language, and one condition varied at a time.

It works in any language pair, and it is built out of the sheets above rather than out of anything we have measured:

1. The name of whoever wrote and checked each language side, because a name is checkable and an assurance is not.
2. A prompt set in two languages, matched on intent, with both wordings written down and the mismatches marked.
3. A name-variant list for your brand in both scripts, published on the site as well as recorded in the set.
4. The same thirteen fields recorded per run, frozen before the first run.
5. The same number of runs per language, so neither side is measured harder than the other.
6. One condition varied at a time, with the interface language and the location recorded every time.
7. A re-test date, and the same questions asked the same way on it.

Where the language decision sits in a testing process

The language decision belongs at the audit step, before anything is run. Lunasol publishes its process in four steps: an audit it calls "The AI census", signals work to "Build what AI trusts", content described as "Write the way AI reads", and quarterly tracking, "Numbers, every quarter" (Lunasol). It describes the tracking step as re-asking "the models your customers' questions" on a schedule. In a process of that shape, the audit step is where a prompt set gets decided, which is where the language question lands, and the tracking step is where a prompt set gets asked again.

Lunasol, which owns AI Visibility, publishes its own GEO and AEO page at two addresses, an English version at lunasol.ae/geo-aeo and an Arabic version at lunasol.ae/ar/geo-aeo. Checked on 21 September 2026, each carried a link to the other, and the process steps were named on both, Audit (التدقيق), Signals (الإشارات), Content (المحتوى) and Track (التتبع) (Lunasol). That is what the two pages showed on that date.

Where to start with your first English and Arabic pair

Run one pair this week in your own sector: the comparison question in English, and the same intent in Arabic, written by somebody who would ask it that way. Record the thirteen fields. One pair will not tell you whether you have a language problem, because nothing this small can, but it will show you what the sheet captures and what a test worth commissioning would have to be built from.

Lunasol, which owns this publication, offers a free AI visibility check as a starting point, if you would rather not build the sheet yourself. Lunasol's published process and its contact routes, WhatsApp and email, are at Lunasol.

Common questions

Should I test AI search in Arabic as well as English? If you sell to Arabic-speaking customers, yes, and as a separate test rather than a translated one. In the single pair we ran, the English and Arabic answers named ten businesses between them and agreed on four.

Is an Arabic version of my website enough for AI search? No. Google determines a page's language from its visible content, so an Arabic menu over English body copy is still an English page. It also asks that each language version be linked to the other in both directions, or the tags are ignored.

Should I write my test questions in Modern Standard Arabic or Gulf dialect? Modern Standard Arabic for the core written set, because dialect has no fixed spelling and Modern Standard Arabic gives you one string you can repeat run after run. Then add Gulf dialect wording as a recorded variant, and Latin-script Arabic if your customers write that way. They are entries on the same sheet, not replacements for each other.

Do I need an Arabic version of my brand name? You need to decide which one you want in circulation and then publish it as text. In the Arabic answer we logged, the assistant transliterated every business name itself and printed the Latin form in brackets beside it, so a name may well exist in Arabic whether or not you chose it.

Does asking in Arabic get a different answer from asking in English? In our one pair, in one tool, on one day, it did: different businesses, a different shape, and a different kind of source. One pair is not a finding about your market, and the study that would be is separate work we have not run.

Will an AI assistant answer my Arabic question in Arabic? In the one Arabic run we logged, in Google's AI Mode, yes: the prose came back in Arabic, while the map card's own groupings rendered in Latin script. That is one run in one tool, and the second arm of our check produced no usable answer, so verify it in your own account rather than assume it, and check the whole answer rather than the first paragraph. That is why the recording sheet has three separate fields: the language of the prompt, the language the answer came back in, and any part of the answer that switched.

Do I need hreflang tags on my Arabic and English pages? Google asks for them if you publish the same content in both languages, and they have to point both ways: "Each language version must list itself as well as all other language versions", and "If two pages don't both point to each other, the tags will be ignored". Google adds that where a complete set becomes difficult to maintain, you can omit some languages on some pages and it "will still process the ones that point to each other" (Google Search Central).

How many test runs do I need in each language? More than one per language, and the same number on each side. A single run tells you what happened once.

What this test cannot tell you

This test cannot tell you which language serves your brand better. Two runs cannot support that and neither can twenty, unless they are run the same way on both sides over time, and we have not done that work. The English against Arabic study is a separate piece of research that has not been run, and nothing in this article is a preview of its results.

It cannot promise you a mention in either language. OpenAI's help pages state that "ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information", that "Placement is not guaranteed", and that "Search results and citations can be incomplete, outdated, or incorrect" (OpenAI Help Center).

And it cannot tell you whether your Arabic reads well. A test records which pages were cited and which names were named. Whether the Arabic on your own site is worth reading is a judgement for a person who writes Arabic for a living, and it is worth paying for separately. Nor can a test tell you which languages a supplier can write and check a prompt set in. That is a question to put to any supplier directly, and this article makes no claim either way about any of them.

Method, sources and notes

This article is general guidance on designing a test, current at September 2026. It is not legal, tax, contractual, data protection or licensing advice, and nothing in it describes what any UAE authority requires or permits you to publish, in either language. Nothing here should be read as a statement about what any trade licence, free zone licence or regulator registration allows. The Arabic and English name forms discussed above are described as spelling conventions only, never as legal designations, and you should check your own trade licence and your regulator's rules before changing, translating or publishing any form of your registered name or its legal suffix. Where we suggest publishing company facts in a second language, that includes facts about individuals: publish a person's name, role, photograph or direct contact details, in either script, only with that person's agreement, because they are that person's personal data and remain your responsibility under UAE data protection law and under any free zone regime that applies to you. No test, and no page written in any language, can secure a mention, a citation, a ranking or a position in any AI product, and nobody can promise one.

The live check above was a method demonstration and is labelled as one throughout: one tool, one day, one connection, one sector, one question pair, one recorded run per language, from an account whose browser and interface language were English. It illustrates what a recording sheet captures.

It is not a study, it is not a finding about any market, and nothing in it is a finding about any particular company. Nothing here is a recommendation of, or a comparison between, any supplier, and we neither endorse nor rank any business named in the answers we logged. We name no company and no website from it. We report counts and structure, and one verbatim question the assistant put back to us, which carries no business name, address or other identifier. The two prompts were frozen before any run, along with the note that they are matched on intent and are not translations of each other. We planned four answers and got two: a second assistant returned a short opening fragment and citation chips on three attempts and would not send on a fourth, and since we established no cause we neither name it nor attribute the failure. The frozen protocol allowed up to three attempts per prompt per tool. We stopped at two per prompt, because the same truncation had by then reproduced across two articles, two sectors and two languages, and further identical attempts would have added nothing. Stopping below the cap is within the frozen rule and we record it here. We did not add extra runs on the tool that was working to make the numbers look better. We counted the English run's external hosts but did not write down which they were, so the count of domains appearing in both languages does not exist and is not reported; that is our own error and it is the reason the recording sheet above says to freeze the fields first. The logged runs are available on request.

Assistant answers vary by account, location, interface language and time, and quoted product documentation changes without notice, so check the current version of any page quoted here before relying on it. Google's help page listing the availability of AI Overviews states which languages and countries are covered; it does not state that answers differ by language, and nothing here should be read as a claim that it does.

The Arabic prompt strings in this article are Modern Standard Arabic and were reviewed for correctness and register before publication. Where Arabic appears inside a quotation from a logged run, it is reproduced exactly as the answer printed it and has not been corrected. This article does not assess any organisation's website against the multilingual guidance quoted above, including any site named in it, and nothing in it is a claim about any organisation's Arabic-language capability, staff or experience.

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. Google Search, Google Search Central, Google Search Help, AI Mode and AI Overviews are products or publications of Google. WhatsApp is a product of Meta. All are trademarks of their respective owners.

AI Visibility is owned by Lunasol.