Open data · method

Gulf AI Citation Index: how it is measured

Each month we ask AI answer engines the same 180 local buying questions for Dubai, Abu Dhabi, Sharjah, Riyadh, Jeddah and Dammam, in English, Modern Standard Arabic and Gulf dialect, through their official web-search APIs. We record which sources they cite, classify them by type and language, and publish the results as open data. This page fixes the method before the first edition, which follows the first full collection.

Updated 25 October 2026 · By AYKA TECHNOLOGIES

Why this index

The language of a question changes which sources an AI engine cites (arXiv:2608.30052), yet most published AI-visibility studies test English questions in Western markets. Gulf businesses need to know which sources the engines rely on when customers ask in Arabic and in dialect: official portals, directories, media, social platforms or business websites. Those sources are where a business should be listed and mentioned.

The questions

• 6 cities: Dubai, Abu Dhabi, Sharjah, Riyadh, Jeddah, Dammam.
• 10 categories: dental clinics, restaurants, cafés, perfume shops, gyms, real estate agencies, digital marketing agencies, car rental, pharmacies, abaya shops.
• 3 versions of each question: English, Modern Standard Arabic, and Gulf dialect (Emirati wording for UAE cities, Saudi wording for Saudi cities). Each version is a separate question, not a translation check.
• That makes 180 questions. They use the words customers type, so some say “best”: that is the customer’s word, not a ranking by AYKA. The question list stays the same from month to month so editions can be compared.

The engines

Questions are sent through the engines’ official APIs with web search switched on: Groq browser search, Gemini with Google Search grounding, OpenAI web search, Perplexity Sonar and Anthropic web search, as far as access allows in a given month. Each edition names the engines and model versions it used. We do not scrape consumer chat apps.

What is recorded

• Engine, model, date and question language.
• Every cited page, and where the API shows it, pages the engine read but did not cite. Being read, cited and used are different steps (arXiv:2604.25707).
• Answers with no web search or no citation are kept and counted. Leaving them out would overstate citation rates (arXiv:2607.14035).
• Each cited site is classified: government, directory (including maps and review sites), marketplace, news and media, social network, video, forum or community, reference such as Wikipedia, academic, business website, other. The language of the cited page is recorded too.

What is published

• The share of answers citing each source type, per engine and per question language, with 95% confidence intervals (Wilson method).
• The most cited domains per language, and how much Arabic and English answers share the same sources.
• From the second edition: which domains entered or left the top 10.
• Files: CSV tables, a README and a citation file, under CC BY 4.0.

Limits

• API answers can differ from what the consumer apps show, which may use location, history and other models.
• Answers vary from run to run, so one month is a snapshot. Trends need several editions.
• The index describes sources, not businesses. Websites of individual businesses are anonymised in the published data; it is not a list of recommended companies.
• AYKA is a GEO agency, and the index helps us advise clients. The method is fixed and published in advance so that anyone can check it.

For journalists and researchers: ask for the first edition or the question list through the contact form. For businesses: we can run the same questions for your category and city and show where you appear.