AI listing descriptions for Dubai property: what works and what to check

What AI is genuinely good at in listing copy

Listing descriptions are a near-perfect job for a language model, for a boring reason: the facts already exist. Bedrooms, size, floor, view, furnishing status, community, price, service charge and handover details sit in your system or in the agent's notes. The slow part is turning that into three readable paragraphs, then doing it again in Arabic, then doing it for forty units that all look alike. AI is fast at that translation from structured facts to prose, and it is patient enough to keep the same tone on the fortieth listing as the first.

Where it helps most in a small or mid-size Dubai brokerage:

• First drafts from unit data. The agent fills in the fields. The AI returns a title and a description in your house style. The agent edits instead of staring at a blank box.
• Consistent tone across agents. Ten agents write ten different ways. A shared set of instructions makes every listing sound like the same brokerage.
• Arabic and English together. The AI drafts both from the same facts, so the two versions do not drift apart.
• Rewrites for different channels. One set of facts becomes a portal description, a short WhatsApp message and an Instagram caption.

The rules you still have to follow

In Dubai, property advertising is tied to a permit system called Trakheesi, an online portal run by the Dubai Land Department (DLD, the Dubai government body that registers property and regulates brokers) that issues advertising permits. Bayut, one of the biggest property portals in the UAE, describes the permit as mandatory for property advertisements in Dubai, and says a portal listing needs the permit number entered on the listing draft before it can go live. For print adverts, Bayut says the permit number must be shown on the advert, which must be in English and Arabic.

I am not going to quote a permit fee or how long a permit lasts here. The sources I found disagree with each other, and the right place to confirm both is the official Trakheesi pages and DLD. The same goes for anything legal about what you may claim in an advert: confirm with the relevant authority.

What matters for AI is simple. The AI does not hold a permit, and it has no way to know whether the permit on file matches the unit, the owner and the agent. That check has to be a rule in your process, not a hope. If the listing also becomes a video, the permit, QR and company-name fields must appear in every cut.

A listing workflow that drafts the copy and checks the permit field

Here is how I would set this up for a brokerage, step by step.

1. Unit data goes in once. The agent enters the facts in one form: community, building, beds, size, view, price, status, and the permit number. This is the single source for everything that follows.
2. The AI drafts English and Arabic. It is told to use only the facts in the form. If a field is empty, it leaves the detail out rather than guessing. This instruction matters more than any other.
3. An automatic checklist runs. Before anyone can approve, the system checks that the permit field is filled, that it is not blank text like "pending", that the price in the copy matches the price in the form, and that the copy does not mention a feature missing from the data (for example "sea view" when the view field says "community").
4. A person approves. The agent, or a listings coordinator, reads both versions, fixes anything off and signs it off.
5. It goes to the portal. Property Finder, a large UAE property portal, and Bayut both let you push listings. How you connect is worth being honest about: Property Finder says its APIs (the technical connections that let your software send listings directly) are for approved partners, who have to go through an application, a security and compliance review and an agreement before getting documentation. So a direct connection is not something you can assume on day one. Check the portal's current partner terms, or use a CRM (customer relationship manager, the system where you keep leads and deals) that already holds that integration.

This is the kind of thing I build for brokerages: a small internal tool that sits between your unit data and the portals, with the drafting and the checklist built in. It is not a product on a shelf, it is wired to how your agents actually work. If you would like to see how workflows like this are put together, my page on AI workflows explains the pattern, and the page on AI for real estate covers the wider picture.

Arabic needs a real reader, not just a translation

AI can write fluent Arabic, and for a standard apartment description it is usually usable. But fluent is not the same as right for your buyer. Arabic readers notice stiff, literal phrasing, and property words (view types, finishing levels, community names) can come out in a way a local reader finds odd. My advice is to have someone on your team who reads Arabic natively approve the first twenty or so listings and note the corrections. Then put those corrections into the AI's instructions as examples. After that the review gets quicker, but it should not disappear.

I also would not promise a fully automatic Arabic chatbot or Arabic reply system from this article. That is a separate subject, and what works there depends on the channel and the volume.

Do it yourself with Claude today

Claude is an AI assistant from Anthropic that you can chat with in a browser. You do not need any system to try this. Paste in something like:

You write listing descriptions for a Dubai brokerage. Use only the facts below. If a fact is missing, do not mention it. Write a 90-word English description in a warm, plain tone and an Arabic version of the same facts. Do not use the words stunning, luxurious or exclusive. Do not invent a view, a floor or a handover date. Facts: 2 bedroom apartment, Dubai Marina, 14th floor, 1,150 sq ft, marina view, furnished, annual service charge confirmed by owner, permit number [enter].

Then ask a second question in the same chat: "List anything in your description that is not in my facts." This catches the quiet inventions. Never paste owner phone numbers, passport details or other private data into any AI tool unless your company policy and the tool's terms allow it.

What still needs a human

• The facts. AI repeats whatever you give it, including a wrong size or a stale price.
• The permit. A person confirms the right permit is attached to the right unit and agent.
• Claims. "Highest ROI", "guaranteed rental return" and similar lines should not appear unless you can prove them. Strip them out.
• Photos and floor plan. The AI cannot judge whether the photos match the words.

Built-in portal tools, generic chat tools or your own workflow

You do not have to start from nothing. Bayut's MyBayut page says its BayutGPT assistant can generate listing descriptions for properties advertised on Bayut, and CASAFARI (a property data platform) has a button that drafts a description from the listing data already in its system. Both are convenient. What neither page tells you is how the text handles a missing fact, which languages it writes, or how it connects to your permit check, so test them on ten of your own units before relying on them. A general chat tool such as Claude gives you full control of the instructions but no connection to your records. A workflow of your own, as described above, is the only route that enforces the permit and price checks for you.

A simple decision rule: under about ten new listings a week, a good shared prompt and a human check is enough. More ready prompts are in ChatGPT and Claude prompts for Dubai real estate agents. Above that, or when several agents write for the same brokerage, the form and the automatic checklist save more time than the drafting does.

Sale, rent and off-plan are three different jobs

One prompt for everything is the most common reason AI copy reads generic. Keep three templates:

• Resale: condition, view, floor, furnishing, service charge, and the one or two features a buyer will actually compare against other units in the building.
• Rental: annual rent, cheque options if you offer them, move-in readiness, and what is included. Never reuse sale wording about "investment".
• Off-plan (sold before the building is finished): developer, payment plan, handover date and what the buyer is paying for today. Give the AI the payment plan terms exactly as the developer states them, and tell it to label any return figure as an estimate. Without that instruction models tend to write returns as if they were promised.

Practitioner guides suggest roughly 100 to 200 words for a standard listing, with the key facts early and two or three real selling points explained as a benefit to the reader. Follow the portal's own field limits where they differ.

Claims AI gets wrong, and where to verify them

Risk in the draft	How to spot it	Verify against
Stale or invented price, service charge or fee	A number in the text that is not in your form	Your own unit record, and DLD for fees
Return or rental yield stated as fact	Words like "will", "guaranteed", "up to X%"	Remove, or label as an estimate with its source
Amenity or view that does not exist	Feature appears in text but not in the data fields	Developer brochure, site visit, photos
Unnatural Arabic property terms	A native reader pauses on a phrase	Native Arabic reviewer, then add the fix to your instructions
Legal or regulatory claim	Any sentence about rules, visas or ownership rights	Delete it, or confirm with the relevant authority

On Arabic, one practitioner guide suggests formal Modern Standard Arabic for portal listings and a lighter Gulf style for social posts. That matches what I see: keep two Arabic instruction sets, not one.

Owner versus agent view

An agent mainly wants speed and a first draft that does not embarrass them. An owner or manager wants consistency and protection: no false claim under the brokerage's name, and a record of who approved what. Design for the owner: log the draft, the person who approved it and the permit number used. The Dubai Land Department has published enforcement news about advertising, which is a good reason to be able to show that trail.

How I would approach this

Start with Claude and your best three agents' existing listings as tone examples. Measure how many minutes a listing takes before and after. If the saving is real and the Arabic holds up, then it is worth turning the process into a proper form with the permit check built in, so the rule is enforced by the tool and not by memory. That is the point where a small custom build starts to pay for itself.