GEO in Dubai: how AI assistants choose who to recommend

AI assistants recommend businesses they can find, read and cross-check. When someone asks ChatGPT, Perplexity, Gemini or Google's AI Overviews for "a Russian-speaking dentist in JLT", the assistant runs a search and reads a handful of pages. It then names the businesses whose facts show up clearly and consistently in more than one place.

GEO, generative engine optimisation, is the work of making that true for your business. Most of it overlaps with classic SEO:

pages a crawler can reach
facts written as text
a consistent business listing
a named author

Nobody can promise that an assistant will name you, because the model, the index and the user's wording all change from week to week. You can control what the assistant finds when it goes looking. For a Dubai-specific action plan, see the GEO guide in the Savage Media journal. This article explains the mechanics behind it.

Two ways an assistant builds an answer

An assistant answers from one of two places.

Memory. The model learned from a snapshot of the web, usually months old. Brands that were mentioned often and consistently across many sites are more likely to be known. A café that opened in Al Quoz this spring probably is not.

Live search. The assistant sends queries to a search index, fetches a few pages and writes an answer, often with links to its sources. Local and commercial questions ("near", "open now", "price", "best for") tend to trigger a search, because memory goes stale fast for that kind of question.

For a business in the UAE, the second path is the one you can influence this month. Memory updates on the model provider's schedule, which you do not control.

What happens during a live search answer

The steps differ between products, but the pattern looks like this:

The assistant rewrites the question into several search queries, often more specific than what the person typed. "Good dentist" can become "Russian-speaking dentist JLT", "dental clinic JLT Friday hours" and similar.
It pulls results from an index. Google's AI Overviews draw on Google Search. Other assistants use their own crawlers, partner indexes or both.
It fetches some of those pages and extracts passages that answer the queries.
It compares sources. A fact that appears on your site, your Google Business Profile and a third-party review carries more weight than a fact found once.
It writes a short answer and names a few businesses.

You can drop out at every step:

If your page is blocked, it never reaches step 3.
If your services live inside an image or a PDF, step 3 extracts little.
If your site says Business Bay and your Google profile says Downtown, step 4 has less reason to trust either.

Signal 1: the assistant can reach and read your pages

Start with access. Each AI company runs named crawlers. Your robots.txt file, your hosting firewall or a security plugin decides which of them get in.

Crawler	Company	What it does
Googlebot	Google	Google Search, including AI Overviews
Google-Extended	Google	Controls use of your content for Gemini models; blocking it does not remove you from Search
Bingbot	Microsoft	Bing index, used by Copilot
OAI-SearchBot	OpenAI	Pages shown in ChatGPT search answers
ChatGPT-User	OpenAI	Fetches a page when a user's request needs it
GPTBot	OpenAI	Collects data for model training
PerplexityBot	Perplexity	Perplexity's search index
ClaudeBot	Anthropic	Collects data for model training

You can let search crawlers in and keep training crawlers out. That is a business decision. Make it on purpose instead of inheriting it from a plugin default.

After access comes readability. Put these in plain HTML text:

services
prices or price ranges
area
hours
languages

A menu that exists only as a PDF, or a price list baked into an Instagram graphic, gives the assistant little to quote.

Signal 2: the same facts everywhere

Assistants cross-check. Pick one version of these facts and use it on every surface:

business name, spelled the same way
area and full address
phone and WhatsApp number
opening hours, including Friday
main services
languages spoken by staff

These surfaces matter most in the UAE:

your website footer and contact page
your Google Business Profile
Apple Maps
Yandex Business and 2GIS, for a Russian-speaking audience

On the site, repeat the same facts in structured data. Use LocalBusiness, Restaurant, MedicalClinic or another fitting type, and use the sameAs field to link your official profiles. The Google side of this is covered step by step in the local SEO and Google Business Profile checklist.

Signal 3: pages that answer the exact question

Assistants rewrite vague questions into specific ones, so specific pages get picked up.

Hypothetical example: a dental clinic in JLT has one "Services" page listing twelve treatments in a single paragraph. A competitor two towers away has a separate page for implants. That page shows:

the price range in AED
the dentist's name and the languages she speaks
the clinic's Friday hours
a short answer to "How many visits does it take?"

When a user asks for "implants in JLT with a Russian-speaking doctor", the second clinic gives the assistant a passage it can lift almost word for word.

Build pages around real combinations of service, area and audience that your business serves. Two projects I built follow this pattern:

The Lucent Rental site has pages for each car, each brand and each brand-by-location combination.
Lucent Legal publishes 213 guides in English and 213 in Russian, each on its own topic.

Inside each page, lead with the answer. Put the price range, the condition or the "yes, we do this" in the first two sentences, then explain.

Signal 4: other sources say the same thing

Your own site is one source. Assistants also cite reviews, directories, local press, comparison articles and forum threads. A business that only describes itself gives the assistant one voice to rely on.

Practical steps:

Ask customers to mention the specific service and area in their Google review ("implant", "JBR", "Russian-speaking"). Specific reviews give the assistant matching text.
Reply to reviews with the same wording you use on the site.
Get listed in the two or three directories your customers use, with the same name and number.
If a local publication writes about you, link to that article from your site so the two sources reinforce each other.

Signal 5: a named author with a real page

An assistant that weighs a medical or legal answer looks for a person behind it. Give it one in four places.

The byline. Every article and expert page shows a full name and a "last updated" date near the top.
The author page. One URL per person, with: - name and role - qualification or licence relevant to the topic - links to official profiles such as LinkedIn - a list of the pages they wrote or reviewed
The markup. The author page carries Person structured data with name, jobTitle and sameAs. Each article carries Article data whose author field points to that page, plus dateModified.
The match. The name is spelled the same on the site, on LinkedIn and on your Google profile, if the person is listed there.

For a clinic, the treatment page names the doctor and shows their licence details. For a law firm, the guide names the lawyer who reviewed it. The Lucent Legal site has a dedicated founder page that its guides can point to. In the Savage Media journal, my articles carry my name and link to an author page.

Test it yourself. Open any article and check two things:

You can reach the author page in one click.
The author page links back to the article.

When you hire a contractor, ask: "Which pages will carry an author, who is it, and where is the Person markup?"

Where llms.txt fits

llms.txt is a proposed format, described at llmstxt.org. It is a Markdown file at the root of your site that lists your key pages with a one-line description of each. It gives a language model a short map of the site without crawling everything.

Adoption by the large assistant providers is unclear, so treat it as a cheap extra. It does not replace crawlable pages or consistent listings. On the sites I build, I add it next to the sitemap. Moon Mars has one live at moonmars.ae, and the Lucent Rental build includes one. The format and a template are in the llms.txt guide.

What the Dubai venue data shows

The bar in Dubai is low, which makes the basics worth doing. The Savage Media study of 537 Dubai restaurant and venue websites, checked on 13 October 2026, found:

55 of 537 sites (10.2%) did not open at all.
52 sites blocked automated visitors, which can also shut out AI crawlers.
Of 409 fully analysed sites, 147 (35.9%) offered the menu only as a PDF.
90 (22.0%) used Restaurant structured data, and 22.7% had no meta description.
Only 12 sites (2.9%) had all five basics: HTTPS, a meta description, a web-page menu, online booking and a WhatsApp button.

A venue that fixes those five points already stands out from most of the sites an assistant would compare it with.

How to check what assistants say about you now

Run a monthly check before you change anything, so you have a baseline:

Write five or six questions your customers ask, with area and language: "Where can I rent a Lamborghini in Dubai Marina for a weekend?", "Russian-speaking lawyer in Dubai for a rental dispute".
Ask them in ChatGPT, Perplexity, Gemini and Google, in a fresh or logged-out session.
Record who gets named and which pages the assistant cites.
Repeat in Russian or Arabic if those are your customers' languages.

The cited pages show you what the assistant trusts in your niche. A fuller method is in how to check what ChatGPT says about your business.

What not to do

Hiding keyword text for bots. This covers white text, collapsed blocks stuffed with area names and a different page for crawlers. Search engines treat this as spam, and assistants inherit that judgement.
Buying or faking reviews. Platforms remove them, and a burst of identical five-star reviews weakens the cross-check you need.
Letting a security plugin decide. Check your firewall and robots.txt before you assume assistants can see you.
Publishing near-duplicate pages. Fifty area pages with the same text and a swapped neighbourhood name add no new facts for an assistant to use.
Rebranding in pieces. Changing your name on Instagram but not on Google Maps splits your business into two entities.
Paying for "guaranteed ChatGPT placement". Ask the seller which crawler, which index and which source they will change, and get that list in writing before you pay.

Checklist

Check	Where	Done when
Search crawlers allowed	`robots.txt`, hosting firewall	Googlebot, Bingbot, OAI-SearchBot and PerplexityBot can fetch your key pages
Facts in plain text	Service, menu and contact pages	Prices, area, hours and languages are readable without opening a PDF or image
One version of your facts	Site, Google Business Profile, Apple Maps, directories	Name, address, phone and hours match character for character
Structured data	Site code	Correct `LocalBusiness` subtype with `sameAs` links to your profiles
One page per real question	Site	Each main service has its own page with the answer in the first two sentences
Named author	Articles, expert pages	Byline, author page with `Person` markup, and last-updated date are visible
Specific reviews	Google Business Profile	Recent reviews mention service and area, and you reply to them
llms.txt and sitemap	Site root	Both files exist and list your current key pages
Monthly baseline	ChatGPT, Perplexity, Gemini, Google	Same questions asked, named businesses and cited sources logged

If the site fails the first three rows, fix that before anything else. That part of website work decides whether an assistant can read you at all. Then run the monthly check again and compare the cited sources with your baseline.

Questions and answers

Can anyone guarantee that ChatGPT will recommend my business?

No. The model, the search index behind it and the way each person words the question all change. A contractor can make your facts easy to find, read and confirm, and that is the part worth paying for.

Is GEO different from SEO?

Most of the work is the same: crawlable pages, clear facts, consistent listings, a named author. GEO adds attention to how an assistant extracts and cross-checks short passages, and to which AI crawlers your site lets in.

Do I need an llms.txt file?

It costs little to add, and it gives a machine-readable map of your key pages. How much assistants use it is unclear, so treat it as an extra on top of a crawlable site, never as a replacement.

How often should I check what assistants say about my business?

Once a month is enough for most small businesses. Use the same five or six questions each time, in a fresh session, and note which sources the assistant cites.