When someone asks an AI assistant "where should I go?", the answer is a handful of names. The Radar measures whose names come up — and, unusually, it measures both surfaces separately: what the model knows on its own, and what it says when it can search the live web. Those are different questions, and conflating them is how most "AI visibility" claims fall apart. Not guessed. Measured, dated, and scoped.
Wave 1 — Dubai (n = 100) · Wave 2 — Riyadh (n = 50) · Wave 3 — Doha (n = 48) · Wave 4 — Kuwait (n = 48) · both surfaces measured · + Dubai F&B pilot (n = 10)The most-reviewed business in the sample — 4.7★ across 2,374 reviews — was never named, on either surface. Nor was the next one at 4.8★ with 2,189. Meanwhile a firm with 34 reviews was named in every single query. Review count and AI visibility are close to unrelated in this data, which is the opposite of what most owners are told.
Of the businesses the model could not name from its own knowledge, 18 in 100 were named once it was allowed to search the live web. Those firms are not unknown — they are simply not remembered. That is a different problem, with a different fix, from the 59% that stayed invisible either way.
Being named by an AI assistant and ranking in the Google Maps top three do not move together in this data. Businesses appear in one without the other in both directions. Optimising only for the map means playing half the game — and the pilot wave showed the same pattern in a completely different sector.
Most claims about AI visibility quietly test one surface and describe the other. Wave 1 measured both, for every business, so the two can be separated instead of averaged. This is the single most useful thing we can contribute to this market, and it is the reason the headline number is 59% rather than 77%.
| What we found | Share of the 100 | What it means for that business |
|---|---|---|
| Never named on either surface | 59% | Neither remembered nor retrievable. The facts about the business are not present, consistent or quotable enough for an assistant to use them. This is the hard case, and the one worth fixing first. |
| Named only when the model could search | 18% | The web knows them; the model does not. They depend entirely on the search layer choosing them at the moment of asking — and lose the entire no-search surface, which is what a raw model call and a browsing-disabled assistant return. |
| Named only from model knowledge | 1% | Rare. Present in training data, yet not selected when live results were available. |
| Named on both surfaces | 22% | Durably visible on the measurement we ran. |
"Is my business visible to AI" has more than one answer, because there is more than one surface. Most published claims about AI visibility never say which one they tested. Ours does — and Wave 1 measured both of them, separately, for every business in the sample.
| Surface | What it is | What a zero there means |
|---|---|---|
| Model knowledge (what the pilot measured) |
The assistant answers from what it absorbed during training, with no live web lookup. This is what you get from a raw model API call, and from assistants when browsing is off. | The model has no durable impression of your business. For a single restaurant that is unsurprising — but it is also the surface where brands become "known" rather than merely retrievable. |
| Search-grounded | The assistant runs a live search and answers from what it retrieves — Google AI Overviews, AI Mode, ChatGPT with search, Perplexity. This is what most consumers actually see today. | You were not retrievable or not selected at the moment of asking. Closer to lost demand, and more sensitive to the work on this site. |
Being explicit about this is the correction we owe our own method. An earlier version of this page described the pilot as measuring what happens "when someone asks ChatGPT, Gemini or Perplexity" — language that points at the search-grounded surface, while the measurement itself ran without search grounding. The numbers are unchanged and the observation stands; the scope claim around them was wider than the evidence. We would flag the same gap in a competitor's data, so we flag it in ours.
Wave 1 delivered on that. Both surfaces were measured per business on 9–10 August 2026, so the difference between "the model doesn't know you" and "the search layer didn't pick you" is now a published number rather than a promise: 18 businesses in 100 fall in that gap. That distinction is, in our view, the most useful thing the Radar can contribute.
Search behaviour is shifting from links to answers. A growing share of "best X near me" decisions — by residents, tourists and corporate buyers — starts inside an AI assistant, and the assistant replies with a handful of names, not a page of results. Global brands have already moved budget to this channel. In the UAE, almost no one is even measuring it yet.
If AI never says your name, you are invisible to that entire slice of demand — and you will not see it in any dashboard you currently use, because nobody's analytics show the recommendation you didn't get.
| Sample | 100 businesses, Dubai — 9 professional-services sectors (dental, aesthetic, clinic, consulting, accounting, real estate, spa, legal, salon) across 8 areas (DIFC, Downtown, Dubai Marina, Jumeirah, Palm Jumeirah, JBR, Business Bay, City Walk) |
| Measured | 9–10 August 2026 |
| AI engine | Google Gemini (gemini-flash-latest), buyer-intent queries per business |
| Surfaces | Both, measured separately for every business. (1) Model knowledge — no search tool attached. (2) Search-grounded — the model was given Google Search and answered from what it retrieved. |
| Queries per business | Up to 5. The count varies by business (5 queries for 63, 4 for 8, 3 for 10, 2 for 13, 1 for 6) because some queries returned no usable answer; those were dropped rather than counted as a zero. "Never named" means zero mentions across every query we could actually measure for that business. |
| Map presence | 8 local buyer queries per business, Google Maps top-3 counted |
| Technical checks | schema.org structured data, AI-crawler access, website availability & response |
| Sample bias — stated deliberately | This is not a random sample of UAE businesses. Every business in it already had a live website, all were commercially established (median 213 Google reviews), and all were drawn from Dubai professional services. The bias runs towards visibility, so the invisibility figures should be read as a floor. A random sample of UAE businesses would almost certainly look worse. |
| Anonymity | Public results are aggregates only; each business's own numbers are shared privately with that business on request |
| Known limits | AI answers vary run to run — results are a dated snapshot of tendency, not a permanent rank. One engine was used, not four. Neither surface is a consumer product surface. And the relationship between structured data and being named remains an untested hypothesis here: we report both figures, we do not claim one causes the other. |
| Sample | 50 businesses, Riyadh — 5 professional-services verticals (aesthetic clinics, dental clinics, law firms, real estate agencies, accounting firms) across 5 commercial districts (Al Olaya, Al Malqa, Hittin, Al Nakheel, As Sahafah) |
| Measured | 1 September 2026 |
| AI engine | Google Gemini (gemini-flash-latest), buyer-intent queries per business |
| Surfaces | Both, measured separately for every business — same method as Wave 1: (1) model knowledge, no search tool; (2) search-grounded with Google Search. |
| Map presence | 8 local buyer queries per business, Google Maps top-3 counted |
| Technical checks | schema.org structured data, AI-crawler access, website availability & response |
| Sampling frame — different from Wave 1, stated deliberately | Wave 1 drew from a curated list where every business already had a live website. Wave 2 sampled what Google Maps actually surfaces to a buyer: the top two results per vertical and district. 42 of the 50 had a live website; headline AI figures are computed over that subset so the two waves compare honestly, and the full-sample split is retained in the raw data. Coverage was complete: all 50 businesses measured on both surfaces, zero lost rows. |
| Anonymity | Public results are aggregates only; each business's own numbers are shared privately with that business on request |
| Known limits | n=50 is a pilot: directional, not conclusive. One engine, one city, five verticals. AI answers vary run to run — results are a dated snapshot of tendency, not a permanent rank. Arabic-language queries were not measured in this wave; buyer queries were English, which likely understates visibility for Arabic-first businesses. |
| Samples | Doha: 48 businesses, 5 professional-services verticals (aesthetic clinics, dental clinics, law firms, real estate agencies, accounting firms) across 8 districts (West Bay, Al Sadd, The Pearl, Lusail, Al Rayyan, Al Waab, Musheireb, Old Airport — 8 because Doha cells run thin). Kuwait: 48 businesses, same verticals across 5 districts (Kuwait City, Salmiya, Hawally, Sharq, Jabriya). |
| Measured | 1 September 2026 |
| AI engine | Google Gemini (gemini-flash-latest), buyer-intent queries per business, English queries |
| Surfaces | Both, measured separately for every business — same method as Waves 1–2. |
| Sampling frame | Same as Wave 2: the top two Google Maps results per vertical and district — what a buyer actually meets — so not pre-filtered to has-website. Headline figures are computed over the live-website subset (Doha 42/48, Kuwait 39/48). Coverage was complete on both surfaces in both cities: zero lost rows. |
| Anonymity | Public results are aggregates only; each business's own numbers are shared privately with that business on request |
| Known limits | n≈50 per city are pilots: directional, not conclusive. One engine. English buyer queries only, which likely understates visibility for Arabic-first businesses — most acutely in Kuwait and Doha where many sampled businesses brand primarily in Arabic. AI answers vary run to run; results are a dated snapshot of tendency, not a permanent rank. |
| Sample | 10 restaurants, Dubai (F&B pilot) |
| Measured | 20 July 2026 |
| AI engine | Google Gemini (gemini-flash-latest), buyer-intent queries per venue |
| Surface | Model knowledge only — no search grounding. The request carried no search tool, so the answers reflect what the model absorbed in training, not a live retrieval of the web. See the surface note above for what this does and does not show. |
| Map presence | 8 local buyer queries per venue, Google Maps top-3 counted |
| Technical checks | schema.org structured data, llms.txt, website availability & response |
| Anonymity | Public results are aggregates only; each venue's own numbers are shared privately with that venue on request |
| Known limits | AI answers vary run to run — results are a dated snapshot of tendency, not a permanent rank. n=10 is a pilot: directional, not conclusive. And the AI-mention figure covers one surface only (model knowledge, no live search); the search-grounded surface is a separate measurement, added in Wave 1. |
What the pilot found about Dubai restaurants · What this means for Dubai clinics and dental practices · Why AI doesn’t recommend your business · Run this measurement on your own business · What is GEO? — Generative Engine Optimization explained · How to appear in ChatGPT recommendations · The 20-point AI visibility checklist · Work with the team behind the Radar · GEO agencies in Dubai, compared honestly · All our services · Local SEO in Dubai — the foundation this measures
AI visibility is whether AI assistants such as ChatGPT, Gemini and Perplexity actually name your business when someone asks them for a recommendation — for example "best Afghan restaurant in Sharjah". It is measurable: ask the engines real buyer questions and count who gets named. It is not the same as ranking on Google Maps, and strong reviews alone do not guarantee it.
Each business is tested against real buyer-intent queries on an AI engine (Wave 1: Google Gemini, 9–10 August 2026, measured on both surfaces and reported separately — the model's own knowledge with no search tool, and search-grounded with Google Search enabled), against 8 local Google Maps queries for top-3 presence, and against technical checks: schema.org structured data, AI-crawler access and website availability. Sample size, measurement date, engine and surface are always published with the numbers, along with the sample's known bias. Public results are aggregated and anonymised; individual businesses only ever receive their own numbers.
mirAIreach runs the same measurement used for the Radar as a free scan for any UAE business — AI mentions, map presence, structured data and website health, scored 0–100 with the weak points identified. Request it via WhatsApp or email; results are private to you.
We run the same measurement used for the Radar on your business: AI mentions, Maps presence, structured data and website health, scored 0–100 with the weak points identified. Private to you. No obligation.