Every paid media plan we inherit from a previous agency in the GCC has the same shape: mostly English-language creative, with the Arabic produced last, in fewer variants, and tested less. The reason given is always some combination of "Arabic is harder," "we don't have the writers," or "the dialect question is too complex."
All three reasons were genuine. None of them are still binding constraints in 2026. The Arabic creative bottleneck was a labour-supply problem, and labour-supply problems dissolve fast in an AI-native operating model.
Why Arabic was hard
Four structural challenges made Arabic creative slow and expensive in a traditional agency model:
Right-to-left layout. Arabic reads RTL, which means every layout, button position, image composition, and visual rhythm has to be rebuilt rather than translated. A Latin-script ad does not become an Arabic ad by swapping the words. Most agencies treated this as a final-step localisation task, which produced layouts that worked grammatically but felt off-balance to native readers.
Dialect variation across the Gulf. Modern Standard Arabic (MSA) is universally readable but emotionally flat. Khaleeji dialect plays well in the UAE, KSA, Kuwait, Bahrain, and Oman with subtle local variations. Egyptian dialect dominates entertainment but reads as out-of-place in a luxury hospitality ad in Riyadh. Levantine has its own tonality. Picking the wrong register is the kind of mistake that's invisible to a non-native marketer and obvious to the audience.
Cultural nuance. Arabic is denser with religious, social, and gendered cues than English. Word choices that are neutral in English carry implication in Arabic. Phrases that work in one Gulf market are awkward in another. Agencies dealt with this through native-writer review, which created a serial bottleneck.
Type and typography. Arabic typography is genuinely different from Latin typography — kerning, ligatures, baseline behaviour, font availability. A Latin font does not have an Arabic counterpart by default; mixing scripts in a single layout requires deliberate font pairing that most design teams don't have a strong opinion on.
Stack those four constraints on top of the standard creative production pipeline and Arabic became the slowest, most expensive, lowest-volume part of every campaign. The result was a chronic under-investment that left brands with weaker performance in their largest market.
What changed
Three things shifted between 2024 and 2026 that, together, dissolved the bottleneck:
Models got materially better at Arabic. The base capability of large language models on Arabic — generation, dialect handling, register sensitivity — improved sharply. The 2026 generation handles Khaleeji, MSA, and Levantine with the kind of register awareness that previously required a seasoned native writer. They aren't perfect, but the failure modes are now checkable rather than starting-from-scratch.
Specialised classifiers became deployable. We can now run a sentence-level dialect classifier and a register classifier on every piece of Arabic creative output. The classifiers were research artifacts in 2023; they're production tools now. Combined, they catch nearly all of the previously-invisible "this would never be said in this dialect" errors before the variant leaves the pipeline.
RTL layout primitives matured. The visual generation models now handle RTL composition natively rather than mirroring a Latin layout. Type rendering with Arabic ligatures works in the variant pipeline. The visual side has caught up with the language side in a way it hadn't a year ago.
What a modern Arabic workflow produces
For a comparable campaign, a modern Arabic workflow produces:
- Several times the Arabic variant volume of a traditional agency baseline.
- Several dialect-and-register variants per audience cohort, where the previous norm was one MSA variant.
- RTL-native layouts, not mirrored Latin layouts.
- Dialect and register checks shown with every variant.
Classifier disagreements are flagged rather than buried, so a brand's attention goes to the highest-judgement work instead of the production grind: steering the dialect mix and choosing archetypes on evidence. The hours once spent writing five MSA headlines go into judgement work that compounds.
What this looks like in performance numbers
The biggest shift we've seen across our Gulf work: Arabic creative win-rates are closing on English ones. Not long ago, Arabic creative trailed English by a wide margin for the same brand. For many brands that gap has narrowed sharply.
Brand-side, this changes the budget allocation conversation. If Arabic-language placements are now performing as well as English ones, the audience-share tilt of the budget should shift to match the audience-share of the market — which in much of the GCC is heavily Arabic-dominant. Most brands we inherit are still allocating based on the legacy creative-volume constraint rather than the current performance data. Reallocating can find incremental return without spending an additional dirham.
What still needs more than a score
Three categories of Arabic creative carry stakes that classifier scores alone cannot settle, however many model layers have scored them:
- Religious and seasonal references. Ramadan, Eid, National Day campaigns. The cost of getting the tone wrong is asymmetric — there's no upside to using a model output on classifier scores alone.
- Comparative or competitive claims in Arabic. Defamation thresholds and the cultural posture toward direct competitor mentions vary across GCC markets. This is judgement work, not pattern-matching work.
- High-spend launch creative. When an archetype is going to receive seven figures in spend across a quarter, a subtle off-brand miss compounding across a million impressions costs far more than the time it takes to get the creative right first time.
Outside those categories, the volume is high, the win-rate is competitive, and the bottleneck is no longer creative production. It's the same bottleneck as in any AI-native engagement: strategic clarity. Which is, on balance, a better problem to have.