Why Arabic and English AI Search Diverge in the Gulf

Arabic and English prompts asked of the same AI engine, on the same day, in the UAE do not reliably pull from the same pool of sources. The clearest documented evidence is a platform-mix inversion inside Google AI Overviews: Instagram supplies just 4.6% of social citations in English-language queries in the Gulf but 29.0% in Arabic ones, while Reddit's share collapses from 21.0% to 4.9% over the same language switch. That is not a translation quirk. It is a structural difference in what these systems consider a trustworthy source, and it has direct consequences for any business trying to be cited in both languages.

What Actually Changes When the Query Language Changes

Profound's March 2026 study, which analysed 3.25 billion Google AI Overview citations, found that query language can "aggressively amplify" social content in some languages and "nearly erase it" in others, with Arabic falling into the second category alongside Swedish. YouTube's share of social citations drops from 38.0% in English to 26.0% in Arabic, still present but no longer the leading platform, while LinkedIn's share roughly doubles, from 5.0% to 8.0% in the Profound dataset (a related figure elsewhere in the same research places the LinkedIn shift closer to 8.0% versus a lower English baseline, depending on the exact query set measured; the direction of the shift, upward, is consistent across the reporting even where the precise magnitude varies slightly by sample).

Crucially, this is a Google AI Overviews-specific pattern. ChatGPT's social sourcing does not adapt the same way by market: its citations remain Reddit-dependent at roughly 51% in the UAE and 57% in the US, because Reddit is embedded in how the model retrieves social content generally, in a way that does not flex per language the way Google's system does. A brand optimising for "AI search" as a single undifferentiated channel is really optimising for several channels that behave nothing alike.

Why the Divergence Happens: Three Separate Mechanisms

The gap is not one phenomenon with one cause. Research reviewed for this piece points to at least three separate, compounding mechanisms.

Tokenisation and training-data scarcity. Arabic accounts for under 1% of total digital content globally, despite being the fourth most-spoken language in the world. Even models purpose-built for Arabic, such as early Jais releases, were trained on datasets containing more than 70% English content specifically to borrow cross-language transfer capability. That imbalance means Arabic queries are answered from a structurally thinner source pool than English ones, before any question of query intent is even considered.

Dialect variation within Arabic itself. Modern Standard Arabic, the formal register used for legal, financial, medical and government content, is not the same as the Gulf Arabic dialects UAE residents use in everyday, consumer-facing queries. The TARJAMAT research assessment of ChatGPT and Bard across ten Arabic varieties found that performance varied meaningfully by dialect, with Modern Standard Arabic consistently outperforming Gulf dialectal variants. GPT-4o handles dialect translation better than smaller Arabic-specific models, but even strong general models show measurably lower accuracy once a prompt shifts from formal MSA into an everyday Gulf register.

A demonstrated, cross-domain performance gap. This is not specific to marketing or commercial queries. A comparative study of ChatGPT-4 and Gemini on identical virology multiple-choice questions found ChatGPT-4 scoring 80% correctness in English versus 65% in Arabic, with Gemini at 62.5% versus 55%. A separate study of infectious-disease questions found "consistently superior performance" in English across every AI model tested. A January 2026 benchmark on Arabic-prompt tool-calling, evaluating five model families including GPT-OSS-20b and several Qwen3 variants, documented a "compounding language penalty": accuracy fell further when both the user's query and the underlying tool description were in Arabic than when an Arabic query was matched against English-language content, evidence that these models reason in a fundamentally English-centric way even when responding in Arabic.

What HAQQ's Legal Retrieval Test Adds to the Picture

Most of the research above measures answer quality or model accuracy. A May 2026 study from HAQQ took a different angle, testing retrieval directly: four matched legal topics, each queried once in English and once in Arabic, run through the Linkup search API. Arabic queries returned nine primary-law sources across the four topics; English queries returned one. The same study documented what it called "jurisdiction contamination," where an Arabic-language UAE labour-law query silently returned Jordanian and Saudi legal sources with no flag distinguishing them from UAE-specific material. HAQQ sells Arabic legal AI tooling, which is a commercial interest worth naming, and the study measures legal primary-source retrieval specifically rather than general commercial recommendations. Even with those caveats, it is a rare piece of controlled, Arabic-specific evidence that retrieval quality, not just answer quality, changes by query language in ways that can misattribute content to the wrong jurisdiction entirely.

Dubai Is Not Really a Two-Language Market

Treating this as a simple Arabic-versus-English binary undersells how fragmented the actual audience is. One useful framing splits Dubai's search population into four practical groups: Emirati users, who move fluidly between Modern Standard Arabic, Gulf dialect, English and code-switched combinations of all three; Arab expatriates, who bring Egyptian, Levantine or other dialectal Arabic into a market where Gulf Arabic is the local default, meaning Gulf dialect should not be assumed to be the only Arabic variant worth testing; South Asian expatriates, who commonly search in English alongside Hindi, Urdu or other South Asian languages, and for whom English commercial content remains essential rather than optional; and international professionals, who search primarily in English using global business terminology. A GEO programme built only for "Arabic" and "English" as two monolithic buckets will still miss the dialectal variation inside the Arabic bucket and the language mix inside the expatriate segments.

What a Real Test Would Actually Need to Look Like

Because no one has published the specific commercial test this cluster keeps flagging as missing, it is worth being precise about what a defensible version of it would require, both to set expectations for anyone trying to interpret a future study and as a genuine methodology a business could commission. A credible design needs at least 60 commercial questions spanning informational, comparative, local and transactional intent, not a handful of similar prompts. Language variants need to cover native English, native Modern Standard Arabic, natural UAE or Gulf Arabic where the query type calls for it, and a set of code-switched prompts reflecting how Emirati and expatriate users actually mix languages in practice. Engine coverage needs to span ChatGPT, Gemini, Google AI Mode and Overviews, Perplexity and Copilot at minimum, plus any Arabic-centric model that is publicly accessible at the time of testing. Controls matter as much as coverage: same location settings, logged-out sessions where possible to avoid personalisation contaminating the result, and multiple runs per prompt-engine-language combination, three to five repetitions being the minimum needed to distinguish a real pattern from ordinary generative variability. The resulting metrics worth tracking go beyond a simple yes-or-no on divergence: brand overlap, defined as the shared set of named vendors divided by the total union of vendors named across both languages, and citation overlap, the same ratio applied to underlying sources rather than vendor names, give a quantified divergence score rather than an impressionistic one. Human review matters too, ideally two native Arabic reviewers plus one English reviewer, with an explicit process for resolving disagreement rather than averaging it away.

Low brand overlap combined with a statistically meaningful difference in local-source share would be the strongest available evidence that Dubai's commercial AI search behaves the way the Bahasa Indonesia test already demonstrated it does elsewhere. Nobody has published that result yet. Running it, transparently, with the methodology disclosed rather than just the headline finding, is one of the more genuinely differentiating pieces of original research a GEO agency operating in this market could produce.

Dialect Variation Goes Deeper Than Gulf-Versus-MSA

The dialect point made earlier understates how granular this gets. Separate dialect-specific evaluation work has found that response quality varies not just between Modern Standard Arabic and "Gulf Arabic" as a single bucket, but across individual dialect groups within that broader category, with Jordanian and Tunisian prompt sets producing measurably different quality outcomes from each other in the same evaluation. That finding matters practically for a UAE-focused content programme in two ways: it confirms that treating "Arabic" as internally uniform, even after correctly separating it from English, still risks flattening real variation the model responds to, and it reinforces why native review needs to come from someone actually familiar with Gulf-specific register rather than a general Arabic speaker from any dialect background, since fluency in Arabic broadly does not guarantee sensitivity to the exact register a Dubai audience expects.

Where Jais 2 and Falcon Change the Calculation

Everything above describes how global models, ChatGPT, Gemini and Google's AI layers, handle the Arabic/English split. Jais 2 and Falcon are a different case entirely, because their training corpus is not a global dataset with Arabic added on top; it is anchored specifically in Gulf-region Arabic content from the outset. That has a practical consequence for the open research question raised below: a commercial vendor-list divergence test run only against ChatGPT or Google AI Overviews would still miss half the picture in a market where Jais 2 and Falcon are positioned as the Arabic-first citation surfaces of choice for exactly the audience segment, Emirati and Arab-national users searching in Arabic, where the divergence matters most. Entity authority, author credibility and structured content need to be established inside the Gulf Arabic web these models draw from, not only on the English-language platforms where most GEO measurement currently focuses.

The Question That Remains Genuinely Open

Here is where honesty matters more than a confident headline. A separate large-scale study, Temso AI's "Lost in Translation," analysed more than 7 million citations across ChatGPT, Copilot, Grok and Google AI Overview specifically on commercially-oriented prompts, the kind a business actually cares about. It confirmed that language reshapes which sources get cited, with local-language citation rates ranging from 85.4% on Google AI Overview down to 51.7% on Grok. Its six tested non-English languages were Spanish, Dutch, German, Swedish, Italian and French. Arabic was not included.

That means no published, controlled study has yet tested whether an Arabic-language commercial query in the UAE, something like "best payroll software for a Dubai company," returns a different named shortlist of vendors than the English equivalent, on the same engine, the same day. A comparable test has already been run and published for Bahasa Indonesia, and it found zero overlapping vendors between the Bahasa and English answers to an identical prompt. Whether Dubai shows the same pattern, a smaller one, or something different again, is an open question. Treating it as already proven would be getting ahead of the evidence; treating it as irrelevant would ignore everything the adjacent research already shows about citation-source and retrieval divergence.

What This Means for How a Bilingual Page Should Be Built

Regardless of how the open question above eventually resolves, the structural response does not change much. Separate, indexable URLs, typically /en/ and /ar/ subdirectories on one domain rather than split across separate domains, with reciprocal hreflang tags on every page and every language variant; partial or one-way hreflang sets are effectively ignored by Google, so both directions need to be present and correct. The Arabic version needs dir="rtl" set at the document root, CSS logical properties such as margin-inline-start instead of hard-coded left or right values, and mirrored navigation and iconography, since a layout that was only ever tested left-to-right tends to break in ways that are easy to miss during development and obvious to an Arabic-reading visitor immediately. Metadata, alt text and schema markup need translating per language rather than left in English on the Arabic page, and Organization, LocalBusiness and FAQPage schema should represent the same real-world entity consistently across both versions rather than describing two different-looking businesses that happen to share a domain.

What Was Tested	Finding	Status
AI Overview social-citation mix	Near-total inversion, English vs Arabic (Profound)	VERIFIED, large-scale
Model correctness, medical/legal domains	Consistently lower in Arabic than English	VERIFIED, multiple studies
Legal primary-source retrieval (HAQQ)	9 sources (Arabic) vs 1 (English), 4 topics	REPORTED, vendor-published
Commercial vendor-list divergence, Arabic vs English, Dubai-specific	No published test exists yet	UNAVAILABLE, open research question

This is one piece of a broader picture. Our complete guide to GEO for Dubai covers the full six-discipline framework this divergence sits inside, and our piece on measuring GEO success in Dubai covers how to track Arabic and English performance separately rather than blending them into one misleading average. For a deeper, platform-by-platform treatment of cross-lingual GEO methodology, our book Cited or Silent devotes a full chapter to the mechanics involved, available as a free gated preview.

---

Frequently Asked Questions

Is Gulf Arabic the same as Modern Standard Arabic for AI search purposes?

No. Modern Standard Arabic, used for formal, legal, financial and government content, consistently outperforms Gulf and other dialectal variants in tested model accuracy. A UAE-facing content strategy typically uses MSA as the primary register for commercial and informational pages, layering in Gulf-specific vocabulary only where query research shows it is actually used.

Does this divergence affect every AI platform equally?

No. Google AI Overviews shows a dramatic, measured shift in its social-citation mix between English and Arabic queries. ChatGPT's social sourcing, by contrast, stays Reddit-dependent regardless of query language. Treating "AI search" as one uniform channel misses this platform-by-platform variation entirely.

Has anyone proven that Arabic and English AI answers name different businesses for the same commercial question in Dubai?

Not yet, specifically for Dubai. The underlying mechanisms, citation-source divergence and lower model accuracy in Arabic, are well documented. A direct, controlled test of commercial vendor-name divergence for the UAE has not been published, unlike an equivalent test that already exists for Bahasa Indonesia.

Is machine translation ever acceptable for Arabic GEO content?

Not as a primary strategy. Every source reviewed for this piece treats native composition, not post-hoc translation, as the baseline requirement, precisely because the performance and retrieval gaps documented above originate in the model's underlying training and reasoning, not just in surface-level wording.

---