How a UAE business gets cited by ChatGPT, Perplexity and Google’s AI answers

For twenty years the job of a website was to be one of ten links, and the job of search marketing was to sit higher on that list than the company next door. The list has not gone anywhere. What changed is that something now sits above it and answers outright, and a large share of people stop reading there.

Three surfaces matter, and they behave differently. Google’s generated summaries, which appear above the organic results for some queries and cite a small number of pages. Assistants people open directly — ChatGPT, Claude, Gemini, Copilot — where there is no results page at all, only an answer. And answer engines such as Perplexity, built around citation from the start.

What they share is the part that decides things for a business: the answer is assembled from a handful of sources, and being one of them is a different achievement from ranking tenth. Tenth place still collects a stray click. Not being cited collects nothing at all.

None of this retires ordinary search work. The pages these systems retrieve are, overwhelmingly, pages that were already findable, already crawlable and already trusted. What it adds is a second question stacked on top of the first — not a replacement for it.

A model answers in one of two ways, and the difference between them is the whole of your leverage.

The first is from memory — what it absorbed while being trained. You cannot edit that. It is frozen at a cutoff date, and for anything local or commercial it is usually stale or simply absent: what a licence costs this year, who actually operates in Ajman, what the going rate is for an app in 2026.

The second is from retrieval. The system runs a search, fetches a handful of pages, reads them, and writes an answer grounded in what it has just read — usually with links back. This is the path that matters to you. Nearly every question a UAE buyer asks about a service, a price or a supplier travels down it, precisely because the model knows it does not know.

So the practical target is narrow and concrete, and it has two halves. Be one of the pages fetched. And be written so that the part of your page which answers the question can be lifted out of it without losing its meaning.

Answer the question in the first two sentences under the heading that asks it. A retrieval system reads a passage, not a whole page. A section headed “What does a website cost in Dubai” that spends four paragraphs on the importance of digital transformation before naming a figure loses to one that names the figure first and explains afterwards.

Be specific enough to be worth lifting. “Prices vary depending on your requirements” is true, useless, and unquotable. “From AED 3,000 for a business site, from AED 8,000 once e-commerce is involved” is a sentence a machine can quote and a reader can act on. Specificity is also self-policing: you can only publish real numbers if you actually have them.

Make each claim stand on its own. Anything that needs the previous three paragraphs to make sense will be skipped or misquoted. Write sentences that survive being extracted alone, and phrase headings as the questions people actually type rather than as clever labels.

Date things, and say who is speaking. “2026 prices” beats “current prices”, because the second sentence is unquotable the moment it is a year old. A page that states when it was written, who wrote it, and what they actually do for a living is easier to trust and easier to attribute — and attribution is the entire thing you are competing for.

These systems are trained on, and retrieve from, a corpus that is overwhelmingly English. Ask an English question about web design in Dubai and there are thousands of candidate pages; the model picks among the most authoritative and you are competing with agencies that have a decade of links.

Ask the same question in Arabic and the candidate pool is dramatically smaller. Worse for the market and better for you, a large part of what does exist is machine-translated from English — thin, generic, and frequently wrong about the local detail that makes the answer useful. A model reaching for an Arabic source on a specific UAE question often has very little to reach for.

That asymmetry is the opening. A genuinely native Arabic page on a specific UAE topic competes for the same retrieval slot against far fewer alternatives, and it reads as the more authoritative of them because it was written by someone who knows both the language and the market. This is not theory: our own Arabic articles reach the first page on queries where their English twins sit on page eight.

It is also the part competitors are slowest to copy, because it cannot be bought cheaply. Running an English page through a translation plugin produces exactly the thin material the model is already choosing between. Writing Arabic content natively is a different exercise with a different cost — which is precisely why so few do it. On keyword research specifically, see why an Arabic site is not an English site translated.

Server-rendered HTML. Many retrieval crawlers execute little or no JavaScript. Content that only exists after a framework boots in the browser may simply not be there when the page is fetched — you will look, to the thing deciding whether to cite you, like a blank page. A static or server-rendered site has no such problem, which is one of the unglamorous reasons we build them that way.

Clean structure. One H1, headings that genuinely describe their section, lists and tables for facts that are enumerable, and no text locked inside images. This is the same advice as ever; it simply matters more when the reader is extracting a passage rather than skimming a layout.

Structured data. Schema.org markup — Organization, LocalBusiness, Service, FAQPage, BlogPosting, Offer — does not compel anything to cite you. What it does is state your facts unambiguously in a format built for machines, so your price, your address and your service area are not being inferred from prose. When the question is “what does this company charge”, having answered it in markup as well as in text removes a guess.

And crawler access is now a decision rather than a default. GPTBot, PerplexityBot, ClaudeBot and Google-Extended are distinct agents you can allow or disallow in robots.txt. Blocking them keeps your content out of training and retrieval; it also removes you from the answers they build. For a business whose actual problem is being found, that trade usually runs the wrong way. There is also llms.txt, a proposed file offering models a clean summary of a site — cheap to add, adoption still unsettled, and no substitute for anything above it.

These systems synthesise consensus. When your business appears consistently across directories, a verified profile, review sites, chambers of commerce and third-party lists — same legal name, same address, same phone number — the model has a stable entity to talk about. When your details differ in six places, it has noise, and noise does not get recommended.

This is the same discipline local search has always demanded, and it now pays twice: once in the map results, once in the answers. If your listing is not verified and complete, start at the Google Business Profile guide before anything on this page.

Recommendation answers in particular lean on third-party lists. When someone asks an assistant for “the best web design companies in Sharjah”, a large part of what it has to work with is other people’s roundups, directory entries and forum threads. Being absent from all of them means the only way you appear is if someone asks for you by name.

Reviews function as evidence rather than decoration. A model summarising what customers say about you needs customers to have said something somewhere. This is one more reason the testimonial and review ask is worth making properly rather than skipping.

Honestly: measurement here is immature, and anyone selling you a rank tracker for AI answers is selling you a sample dressed as a ranking. The same question asked twice can be answered from different sources. Treat the numbers as directional.

What does work is a fixed prompt set. Write down twenty to thirty questions a real buyer would ask — “how much does a website cost in Dubai”, “best app developers in Abu Dhabi”, and the Arabic equivalents, which are not translations of the English ones. Run them monthly across the assistants your customers actually use, in both languages, and record whether you appear and, when you do not, who does. It is manual, it takes an hour, and it is the only direct evidence available.

Then watch referrals. Visits arriving from assistant domains show up in analytics like any other referrer. The volumes are small today; the trend is what you are reading, not the total.

And keep reading Search Console, which still tells you things nothing else does. Impressions holding steady while click-through falls on informational queries is what this shift looks like from the inside: you are still being shown, and the answer above you is absorbing the click.

Expect slowness. Retrieval indexes refresh on their own schedule and training data on a far slower one. A page published this month is not going to be the consensus answer next week, in either language.

Do not publish machine-generated content at volume in the hope of feeding the models. The property being selected for is specificity and first-hand knowledge; bulk generated text is the single most abundant thing on the web and therefore the least likely to be chosen. It also degrades the pages around it.

Do not spin up a page for every service crossed with every city. It reads as a doorway pattern to search engines and as filler to a retrieval system, and it is the exact mistake that buried a sister site under thousands of rejected URLs. Seven substantial city pages beat forty-two thin ones, in both races.

Do not invent statistics to sound authoritative. A fabricated number is the easiest claim in the world to check, and the moment one is caught the rest of the page is worth nothing. If you do not have the data, write the sentence you can actually stand behind.

Do not block the assistant crawlers and then wonder why you are not in the answers. If there is a genuine reason to block them — and for some businesses there is — accept the consequence rather than expecting both.

And do not treat this as a separate discipline with a separate budget. There is no page that is invisible in search and citable by a model. The work is the ordinary work, done well, written so a machine can quote it, and done in Arabic where almost nobody else is bothering.