Measuring AI visibility in the Gulf, the Levant and Egypt: languages, cities and the four metrics

The unit: one answer, in one language, for one place

Every number below is built from single answers: one platform answering one question, asked in one language, for one location, on one day. “Best IVF clinic in Dubai” asked in English and «أفضل مركز أطفال أنابيب في دبي» asked in Arabic are two questions, not one, and each needs its own answers before you can say anything about either.

Mention rate: are you in the answer?

The share of answers that name you. In the example, DAMAC was named in 51 of 120 answers to questions that did not include its name: a 43% mention rate, against 55% for Emaar over the same answers.

Two regional traps. First, questions that already contain your name inflate the rate; keep them in a separate set. Second, matching: Arabic attaches prefixes to names (و, ب, ل, ك), and a business may be spelled several ways in Arabic and in Latin script. A mention written «وداماك» has to count as you, and a different company with a similar name must not.

Rank: where you sit in the list

When an answer lists options, your position in it. The example’s average was #3.2 across the 50 answers that named the brand with a position. Rank exists only when you are named, so read it next to mention rate: a business named rarely but always first can show a better average than one named in every answer in third place. Arabic answers written as flowing prose often imply no order at all; count rank only where the answer really ranks.

Sentiment and framing

Whether an answer presents you positively, neutrally or with a caveat, and how: luxury or value, reliable handover or delays, family-friendly or business-focused. The example scored 68 out of 100. Sentiment is judged by a model reading the answer, so treat it as a direction across many answers and read the answers behind a surprising move, in both languages, since a caveat can appear in one and not the other.

Citations: which regional sites carry the answer

The sources an answer lists. Here they are the most practical metric, because the regional sources are few and specific. In the example, bayut.com was cited 74 times across the 150 answers. For a restaurant group the equivalent might be Zomato, Talabat or Time Out Dubai; for a hospital, health directories and newspapers such as Khaleej Times, Gulf News or The National. A source cited in answers that name your competitors but not you is a gap you can act on.

Platforms show sources differently, and Arabic answers often cite a different set from English ones, so compare citation counts within a platform and a language rather than across them.

Keep Arabic and English apart

Report each metric per language before you report a total. In the example, the English question “best off-plan developer in Dubai” and the Gulf-dialect Arabic question «منو أكثر مطور عقاري موثوق في الإمارات؟» each named DAMAC on three of five platforms, but not the same three: ChatGPT named it in English and not in Arabic, Gemini the other way round. They are different questions, so this does not isolate the language effect, but it shows why one blended figure would hide where the gaps are.

Dialect is a choice to make deliberately. Questions phrased the way customers type, in Gulf, Levantine or Egyptian wording, are closer to real behaviour; Modern Standard Arabic is easier to keep stable across markets. Many businesses track a core set in Modern Standard Arabic and a smaller set in the dialect of their largest market.

Cities and how the platform heard them

Dubai and Sharjah, Riyadh and Jeddah, Kuwait City and Doha can each return their own shortlist, so a business with branches should measure each city separately. How the platform was told the location matters: a structured location sent with the request is closest to a customer standing in that city; a place name written into the question only tells the platform what is being asked about.

Availability matters too. Google AI Mode has reached countries and languages on different dates, so check whether it answers in your market before reading a gap. An answer that could not be measured is “not measured”, never a zero.

How many answers each slice needs

Slicing shrinks samples fast. Twenty questions on five platforms produce 100 answers per run. Split them across two languages and two cities, and each platform sees five answers per language and city. At that size one extra mention moves a rate by 20 points. Before reacting, check the count behind the number and whether the change holds across more than one run.

Your working week and the regional calendar

Compare like periods. Most of the region works Sunday to Thursday; the UAE and Lebanon work Monday to Friday. A weekly report that closes on the wrong day splits your week in two. Ramadan, Eid and the summer season change what people ask and which pages are fresh, so compare Ramadan with the previous Ramadan, not with the month before it, and check whether competitors moved too before crediting or blaming your own work.

How Pharos helps

Pharos measures all four metrics on ChatGPT, Gemini, Perplexity, Claude and Google AI Mode, in English and Arabic, for each city you track. It matches Arabic spellings and attached prefixes with guards against false matches, notes how each platform received each location, keeps every answer to read, and sends your weekly report on the first day of your working week. Measured answers per month are questions × 5 platforms × runs per week × 4.3.