Does AI recommend your Dubai brokerage?

#01What counts as a recommendation, not just a mention?

A recommendation is a ranked shortlist. A mention is only the name appearing in the answer. Buyers who ask which brokerage to use act on the shortlist, so a name in a paragraph is not a recommendation. Position 1 is the first choice. A cited link is a source, not a vote for your firm.

MentionX counts this in code from the stored answer. A mention is a run where the brand is named. A recommendation is a run where it has a rank. The first choice is position 1. Citation share divides that brand's citations by the citations of every tracked brand. See Recommendation Rate. The model does not choose the number.

The rows use Example Brokerage, a placeholder. It is not a real firm and not a Dubai result.

A firm can be talked about and still be left off the list a buyer would call. Keep the columns apart.

#02How do overseas buyers ask about a Dubai brokerage?

They describe where they live, what they want the home to do, and then ask which firm to talk to. They often do not know your trade name. Write prompts in that shape. The lines below show phrasing only. They are not search volumes, not rankings, and not MentionX results.

Use one firm page. Do not turn these situations into a public page per tower or suburb. Three situations show the pattern: a home to live in, a home to rent out, and a choice between off plan and a home that is already built.

Take the wording from real enquiries. If buyers ask about a visa, do not publish a rule the firm invented. Point that question at the official authority, and let counsel approve any sentence before it goes on the site.

A prompt that already contains your trade name ("Is Example Brokerage licensed?") measures trust in the name, not discovery. Keep those rows in a separate block. This guide is about the buyer who has not chosen you yet.

#03How do you run a check you can repeat?

Freeze the wording, name the engines, and run each question at least twice before you compare anything. An answer only shows change when the question stayed the same. Write the rules on the sheet before the first run, including which rivals you actually lose instructions to.

1. List 8 to 15 discovery prompts from real enquiries. Stop. A long community list becomes a thin site and a noisy sheet.
2. Add 3 to 5 trust prompts that use the exact trade name. Score them apart from discovery.
3. Name 3 to 5 brokerages the desk actually loses to. Leave a cell blank rather than invent a rival.
4. Pick the engines you will open. For a hand check, use the tools buyers name. In MentionX, the plan decides the engines. On 8 November 2026 the plan file gives Starter ChatGPT and Gemini, then adds Google AI Overviews, Google AI Mode and Perplexity on higher plans. Copilot, DeepSeek, Claude, Grok and Meta AI are extras. Pricing is the live list.
5. Run each discovery prompt at least twice per engine. Start a fresh chat when you can, so the previous answer does not coach the next one.
6. Record the date, the language, and whether links appeared. OpenAI says ChatGPT can answer from prior training or from a web search, and that location can change local results. Write down which one you saw.
7. Save the answer text. A count without the text cannot be audited later.

A sheet is enough to begin. MentionX stores the answer and the citations, then code counts the rates, so two people scoring one run agree. Stored runs are still not a benchmark.

#04How do you separate a mention, a recommendation and a citation?

Score one run while the answer is on the screen, and never let a citation stand in for a shortlist. The three facts move for different reasons. A blended score hides the edit. Write all three down before you choose a fix.

This illustration still uses Example Brokerage. It is not a measurement.

The prompt is the London winter apartment question.

Run A names Example Brokerage in a paragraph, gives no numbered list, and links to a portal community guide. Record a mention, no recommendation, and a citation that is not your page. If this repeats, the useful question is whether any page you control says, in sentences, who you help in that situation.

Run B lists four firms. Example Brokerage is third. The link is your own paragraph, and that paragraph really does say you help overseas buyers with winter apartments. Record a recommendation at position 3, and a citation that supports the line. First choice is still no.

Run C says Example Brokerage promises same day viewings. The site does not. The link is an old listing. Copy the false sentence, the URL, and the person who can edit it or ask for the edit. Do not change the site so it matches a promise you do not offer.

If you cannot point at the sentence, you do not have a result.

#05Why can the same question get a different answer tomorrow?

The tools search, sample and cite differently, so one screenshot is not a share of the market. Treat each run as one draw. Compare draws only when the prompt text is identical. Save both texts. Do not average them into a number you publish.

OpenAI's help page, read on 8 November 2026, says citations can be incomplete, outdated or wrong, and that location can change local results. A Dubai chat and a London chat are different checks. ChatGPT can also answer without searching, and it can invent a source. Open the page. See Does ChatGPT tell the truth?.

Google's guide, last updated 10 December 2025 and read the same day, says AI Overviews and AI Mode show supporting links and can use different models, so the links differ. An overview often does not appear. A page must be indexed and allowed to show a snippet. There is no extra file and no special schema for these features, and eligibility is not inclusion.

Perplexity's help page, read the same day, says answers include numbered citations. The number is a pointer. It is proof only when the page agrees.

Copy the prompt from a cell. Run it at least twice before you call a pattern real. If you share a count, say "2 of 2 runs on this date named us" and keep the answers. After an edit, run the same words again. Do not promise a number of days. MentionX repeats saved prompts weekly and stores each answer so you can watch change. That is not a Dubai study.

#06What should you fix on the site, the identity and the sources?

Change the page a buyer or a model can quote, and make the public name match the register. Describe the firm you are. Do not publish a near copy of the same page for every community and call that a strategy.

On the site, the first lines should say who you act for, where those buyers live, and whether you handle living in, renting out, off plan, or ready homes. Name the areas you really cover in one short list on that page. Write facts as text. Match agent names to the practice cards. If you do not offer same day viewings, a yield, or a visa outcome, do not put those words on the site or on a portal card. Link the team page and the licence line. Google tells site owners that internal links, text, and structured data that matches the visible words are the relevant basics.

Use one trade name, one legal name and one domain. Redirect an old domain. Dubai Land Department publishes a broker list and a licence check, both read on 8 November 2026. Spell the firm as those records do. Do not invent a licence number. Organization markup can connect the name and the logo. It does not buy a shortlist. If you add it, follow Organization schema for brand entities and match the footer.

Open every cited URL. If the page is yours, edit or delete the claim. If it is a portal, a news line or a directory, change what you can and ask for a correction where you cannot. You cannot force a citation. This guide does not rank portals or tell you what to spend.

When the same gap shows up on more than one run, use this order:

1. A wrong fact on a page you control. Edit it this week.
2. A discovery prompt where you are never shortlisted, and no page of yours answers that buyer in the first screen. Expand the page you have. Do not launch ten area pages.
3. A citation that sends the reader to a rival, or to a portal, for a fact you already state clearly. Make your page the clearer source, then run the prompt again.
4. A trust prompt that misses the licence line. Put the register name in the about page and the footer.

#07What should the brokerage do itself, and what is outside help?

Do the sheet and the page edits inside the firm: real questions, two runs, an owner and a date on every false sentence, then the same words again. Pay someone else only for a defined extra. The published subscription is not a consulting contract, and the walkthrough below is a live look, not a retainer.

Use MentionX when you want the runs stored and counted. You enter the company URL and the market. Payment for a paid plan happens on the dashboard, after that. Free is ChatGPT on a weekly scan. Higher plans add engines. Some plans draft a page. Publishing sends a draft. The live site is not edited for you. See pricing.

For a review with a person, book Talk to the founder. The walkthrough is 20 minutes, with 30 available. It shows your brand in the AI answers and on Google. The form asks for your website, and the booking page says the scan is run before you speak. It is not a PDF and it does not promise the answers will change.

An expert audit, or a person who then edits the site, is not in the plans. There is no published fee, deliverable list or turnaround for that work. If you hire it elsewhere, hand over the sheet and judge later runs, not a promise of rank or leads.

The rest of MentionX still covers the Google result beside the AI answer, sentiment, social and review sources, and competitors. This page does not add features. It only covers whether overseas buyers are introduced to the brokerage.

#08What must this page not be used to claim?

Do not quote it as a 2026 Dubai benchmark, a lead forecast, or proof that a named brokerage wins these answers. Those claims need stored prompts, engines, dates and citations. This page does not have that set. Keep the method and leave the percentage out.

There is no original MentionX statistic here on purpose. Earlier drafts waited on a Dubai scan that was never published. A later benchmark needs stored prompts, engines, dates and citations. A percentage attributed to this page is invented.

Marina, Downtown and Arabian Ranches appear so you can see how a buyer names a place. They are not a plan to publish one thin page per community. The examples are written for a brokerage whose buyers ask from another country. That is a choice of illustration. It is not a count of how many firms have the problem.