GoldieBench Blog · 6 min read
AI SEO Agency Dubai Work: Which AI Models Draft Pages, Schema And Briefs Best, Using Live GoldieBench Scores
GoldieBench has no Arabic or SEO test, so its three Pages tasks are the nearest guide to which models draft Dubai service pages; the rest needs people.

An AI SEO agency Dubai firms hire will use AI models to draft service pages, schema and content briefs, so the useful question is which models build pages well and which Dubai tasks still need people.
Every score on this page was read on live GoldieBench pages on 11 November 2026, and each one names the page it came from.
GoldieBench has no Arabic test, so the language half of the job is covered separately below.
Want people to look at your site too? Goldie Agency offers a free strategy session and works with Dubai businesses remotely.
What GoldieBench tests, and the gap for Arabic
The GoldieBench home page shows 1310 live demos across 50 tasks and 33 models, with Julian's 0–10 scores averaged across every task a model ran.
Tasks sit in four groups: Games, Visuals, Sims and Pages, and Pages holds three tasks.
Nothing on the bench tests SEO writing, schema output, hreflang or Arabic copy.
So the Pages tasks are the closest proxy for the build side of AI SEO, and everything language-related needs a human check.
Landing: the closest proxy for a Dubai service page
The Landing task page asks for a modern marketing landing page in one shot, with 21 of 26 attempts scored and an average of 8.11/10.
| Model | Landing score | Page read |
|---|---|---|
| Fugu Ultra (Sakana AI) | 9.0/10, named winner | /tasks/landing |
| Fusion (OpenRouter) | 9.0/10 | /tasks/landing |
| GLM-5.2 (Zhipu / Z.ai) | 9.0/10 | /tasks/landing |
| Grok (xAI) | 9.0/10 | /tasks/landing |
| GPT-5.6 Sol (OpenAI) | 8.7/10 | /tasks/landing |
| Gemini 3.6 Flash (Google) | 8.6/10 | /tasks/landing |
| Claude Fable 5 (Anthropic) | 7.2/10 | /tasks/landing |
| Kimi K2.7 (Moonshot AI) | 6.5/10 | /tasks/landing |
A high Landing score means a polished layout from one prompt.
It says nothing about whether the words on that page will be cited by ChatGPT or Google's AI Overviews.
Webos and Aipbpromo: the other two Pages tests
The Webos task page averages 7.98/10, with Fusion and Grok both on 9.0/10, GPT-5.6 Sol on 8.4/10, and Claude Fable 5, Gemini 3.6 Flash and Kimi K2.7 each on 8.0/10.
The Aipbpromo task page averages 7.48/10 and names Qwen 3.8 the winner, with GPT-5.6 Sol on 8.3/10 and Gemini 3.6 Flash on 7.8/10.
Across all three, Fusion, Grok and GPT-5.6 Sol stay near the top, which makes them steady picks for drafting page layouts.
The overall leaderboard in context
On the home page leaderboard on 11 November 2026, Fusion led at 8.59, Claude Opus 5 was second at 8.27, Hermes MoA third at 8.17 and GPT-5.6 Sol fourth at 8.16.
GPT-5.6 Sol carries a "best at Pages" tag on that table, and so does GLM-5.2 further down.
The overall average mixes games and simulations with page builds, so for SEO work, weight the Pages scores above the headline number.
Matching models to the Dubai AI SEO tasks
| Task | Can a model help? | What the bench tells you |
|---|---|---|
| English service page layout | Yes, draft | Landing and Webos scores are the nearest guide |
| Arabic service page | Draft only, then native review | No Arabic task on the bench |
| hreflang tag sets for /en/ and /ar/ | Yes, draft and audit | Not tested; check against Google's rules |
| LocalBusiness JSON-LD | Yes, draft | Not tested; check against Google's schema docs |
| Content brief from customer questions | Yes, draft | Not tested; judge on your own data |
| Business Profile edits and reviews | No | Owner's job under Google's policies |
| Outreach for mentions and links | Research only | Relationships need people |
The bilingual rules no model score covers
Google's localized versions guide says each version must list itself and every other version, using fully qualified URLs and valid ISO codes such as ar-AE and en-AE.
Its multilingual sites page says Google detects language from the visible text, advises against side-by-side translations and warns against redirecting people by IP.
A model can produce those tags in seconds, but a person has to check that the pairs actually point both ways on the live site.
For a Dubai clinic, that might mean: draft the English page with a Pages leader, write the Arabic page from its own outline, have a native speaker review it, then add both hreflang sets.
What the platforms say the page needs
Google's AI optimisation guide says optimising for generative AI search is still SEO, and that Search ignores llms.txt and content "chunking".
Its AI features page says a page only needs to be indexed and snippet-eligible to appear in AI Overviews or AI Mode, with no special schema required.
OpenAI's crawler docs say a site that blocks OAI-SearchBot won't appear in ChatGPT search answers.
OpenAI's Business Hub page, checked in November 2026, only lets local shops and restaurants claim a U.S. listing that already exists, so Dubai shops can't lean on a claimed ChatGPT profile yet.
Bing's Webmaster Guidelines ask for consistent entity names across the web, which matters when your brand has an Arabic and an English form.
The human and agency part
Models don't hold your Google Business Profile, earn editorial mentions or know which questions your customers ask on WhatsApp.
Google's Business Profile guidelines rule out repeating the same name in two scripts and listing a virtual office, and only the owner should make those edits.
Goldie Agency covers this side for Dubai clients remotely, with what it calls "manual outreach, no PBNs" for links.
Julian's about page reports 425K+ YouTube subscribers, and his own figures for the agency are 1,000+ campaigns and 240+ clients.
To build the skills in-house first, the AI Profit Boardroom lets you learn AI SEO yourself alongside other owners.
Google is clear that no one can guarantee a #1 ranking, and OpenAI says placement isn't guaranteed, so judge any team on process, not promises.
How to pick a model this week
- Pick one Pages leader, such as Fusion, Grok or GPT-5.6 Sol, for English page drafts.
- Use the same model to draft hreflang sets and LocalBusiness JSON-LD, then check both against Google's docs.
- Draft Arabic from its own outline and send it to a native speaker.
- Confirm OAI-SearchBot, PerplexityBot and Claude-SearchBot are allowed in robots.txt.
- Log each page you ship so you can read Search Console against it.
If you'd rather learn the full workflow, the AI Profit Boardroom is where to learn AI SEO yourself step by step.
When the drafts are ready and you want people on the rest, book a free SEO strategy session.
Two free extras: the 200+ AI SEO Prompts for drafting with any model, and Julian's free link building book for the outreach half.
Also On Our Network
- Agent Operatorsthe nine-step operator checklist for a bilingual Dubai AI SEO engagement
- AI Income Deskwhether a Dubai small business should hire for AI SEO, with three worked examples
- AI Tool Verdictfive ways to get AI SEO done in Dubai, scored out of 10
- agentos.guidefive Agent OS prompts that audit a bilingual Dubai site before you hire
Pick one Pages leader, check your crawler access, and book a free strategy session to plan the human side with an AI SEO agency Dubai firms can work with remotely.
FAQ
Does GoldieBench test Arabic writing?
No. Its task groups are Games, Visuals, Sims and Pages, so Arabic copy needs a native speaker to review it.
Which models scored 9.0 on the Landing task?
On 11 November 2026 the Landing page showed Fugu Ultra, Fusion, GLM-5.2 and Grok at 9.0/10, with GPT-5.6 Sol at 8.7/10.
Will a top-scoring model get my Dubai business cited in ChatGPT?
No score can do that. OpenAI says placement isn't guaranteed, and citations depend on crawler access, content and what the web says about you.
Can a model write my hreflang tags?
Yes, as a draft. Check that every version lists itself and its partner with valid codes such as ar-AE and en-AE, as Google's guide requires.


