How AI assistants decide which businesses to recommend in the Gulf, the Levant and Egypt

Memory or search: the first fork in the road

An assistant can answer from what its model absorbed during training, or it can search the web first and write from the pages it finds. Training knowledge is broad but frozen at a cut-off date, and it favours names that were written about heavily in English: Dubai’s master developers, the big hospital groups, international hotel chains. A new clinic in Salmiya or a family-run bakery in Jabal Amman is unlikely to be there at all.

Questions about local services, prices, current projects or opening hours usually push the assistant to search. That is where a regional business has room to move, because the pages it reads are pages that exist today and can be improved.

The regional web the assistant reads

When an assistant searches a question about this region, it tends to land on a recognisable set of sites. These are patterns seen in answers, not a published list, and they shift by platform, city and sector:

• Property portals and classifieds such as Bayut, Property Finder and dubizzle, for developers, brokers, communities and rents in the UAE.
• English-language newspapers such as Khaleej Times, Gulf News and The National, for launches, rankings, awards and anything that reads as news.
• Lifestyle guides such as Time Out Dubai, for restaurants, hotels, spas, brunches and “best of” lists.
• Food and delivery platforms such as Zomato and Talabat, for cuisine, price level, areas served and reviews.
• Broker, agency and comparison blogs, which publish “top developers” or “best hospitals” roundups that assistants quote freely.
• Map listings, which carry the address, hours and reviews for almost every clinic, shop and restaurant.

Behind the scenes, a single question is often split into several searches. Google calls this “query fan-out” in AI Mode. A question about a maternity hospital in Jeddah might become searches for maternity hospitals, for private hospitals in Jeddah, and for reviews of a few named ones. The businesses that surface across several of those searches are the ones that reach the answer. What AI platforms cite in the region goes through each source type and how to get included.

One question, five platforms, three different shortlists

Claude named DAMAC second, Google AI Mode fourth and ChatGPT fifth. Gemini and Perplexity did not name it at all; both put Emaar first and also named Sobha and Nakheel. Not one of the five answers cited damacproperties.com. They were built from broker sites, agency blogs and portals, and Perplexity’s ten sources included Bayut.

The lesson is not about one developer. A large, well-known business can still be absent from two of five answers to the most obvious question in its category, and what decides its place is mostly what other sites say, not its own website.

How the five platforms differ for a business here

• ChatGPT chooses per question whether to search. When it does, it combines outside search providers with OpenAI’s own crawler and may use the user’s approximate location, so the same question from Dubai and from Cairo can come back different.
• Gemini can ground answers in Google Search, which brings in the Google Business Profile details that many customers here already check on Google Maps before visiting a clinic, shop or restaurant.
• Google AI Mode runs on Google’s index but is not offered everywhere or in every language on the same date. Whether it answers in your country, and in Arabic, is something to check rather than assume.
• Perplexity searches almost every time and lists its sources beside the answer, which makes it the easiest place to see which regional sites carry weight in your category.
• Claude uses web search when it is available and switched on; otherwise it answers from training, where well-documented regional names have the advantage.

None of the five publishes how it picks businesses, and all of them change often. Treat these descriptions as working assumptions to check against real answers.

Arabic questions can reach a different shortlist

A question asked in Arabic sends the assistant to Arabic pages, and the Arabic web about a category is often thinner and made of different sites than the English one. The businesses named can change with it. A clinic with a detailed English site and a one-line Arabic page may do well in English answers and disappear from Arabic ones.

Dialect matters too. Customers rarely type Modern Standard Arabic into a chat box. They write the way they speak: «منو أفضل…» in the Gulf, «مين أحسن…» in the Levant, «فين أحسن…» in Egypt. Assistants understand all of these, but the wording steers which pages match. Writing your own pages in clear Modern Standard Arabic, with the everyday words your customers use for your services, covers both. Arabic vs English explains when tracking both languages is worth it.

Dubai is not Sharjah, and Riyadh is not Jeddah

Recommendations for local services are tied to place. A furniture retailer strong in Dubai may not be named for the same question in Sharjah, and a hospital group known in Riyadh may not come up for Jeddah. Neighbourhoods matter as well: “near Dubai Marina”, “in Al Olaya” and “in Heliopolis” narrow the answer further.

If you serve several cities, each needs its own evidence: a branch page with the real address, a map listing, and mentions in local sources. One national page listing ten city names is rarely enough for an assistant to connect you with any of them.

What makes a business here easy to recommend

The general rule is the same everywhere: be readable, be specific, be consistent and be mentioned by others. What that means in practice depends on the sector.

• Property developers: delivered projects with handover dates, current launches, payment plans and communities, stated on your site and matching what portals and brokers say.
• Hospitals and clinics: specialties, doctors, languages spoken, insurance accepted and the licence your health authority issued, on a page per branch.
• Hotels: location relative to landmarks, room types, family or business focus, and coverage in the guides and newspapers readers trust.
• Retailers and restaurants: complete profiles on the delivery and map platforms your customers use, with areas served, hours (including Ramadan hours) and real reviews.

Why a single answer proves little

Answers are generated fresh each time. The searches, the pages read, the language, the city, the conversation and, on some platforms, the user’s saved memory all nudge the result. Screenshots passed around a marketing team show what one person saw once. Whether you are usually named is a rate, measured over many asks, per platform, per language and per city. Measuring AI visibility shows how.

Nobody can buy a guaranteed place in an organic answer or edit a model’s training data, whatever a vendor promises. What you can change is what the assistant finds when it searches: your own pages, your listings, and what the region’s portals, newspapers and guides say about you. The next guide, a practical checklist for ChatGPT, turns that into steps.

How Pharos helps

Pharos asks your customers’ questions on ChatGPT, Gemini, Perplexity, Claude and Google AI Mode, in English, in Arabic or both, for each city you track, and keeps every answer. You see how often you are named, who is named instead, how you are described and which regional sources the answers lean on. The free audit covers the UAE, Saudi Arabia, Qatar, Kuwait, Bahrain, Oman, Jordan, Lebanon and Egypt.