We Audited 5 Dubai Property Websites for Agent Readiness. None Are Optimized for the Era That Just Arrived.

Last week we published a piece arguing that the agentic web shift demands a different kind of CMS architecture, and that businesses on the wrong substrate will quietly disappear from a class of traffic that didn't exist two years ago. The natural follow-up question: how does Dubai's actual property market look right now against that bar?

We picked five sites — anonymized but representative of the market — and checked each one against the structural items every agent-ready website needs. The headline finding is short: none of them have any of the items. Not because the operators are negligent. Because the entire industry is still building for the previous era.

Why this audit, now

Three things shipped in the last 30 days that change what a website needs to be: Lighthouse 13.3 made agentic-browsing audits part of its default configuration, Chrome 149 opened the WebMCP origin trial, and Google rolled out Information Agents to a Search audience now exceeding 1 billion monthly users. We covered the full thesis in Agentic Search Is Here. Your CMS Decides If You Survive It. This piece is the operational follow-up — what those changes look like on actual websites in the actual market we serve.

We narrowed the sample to property because Dubai's premium property economy is exactly the kind of high-research, high-ticket category where AI Mode citations and agent-mediated discovery will reshape lead flow first. Buyers research for weeks before contacting an agent. The websites they're researching are about to become the substrate that AI agents read, summarize, and act on. That changes the bar.

What we checked

Five items, each observable from a browser or a single curl request. No invented scores. No black-box rubric. Just whether something is structurally present or not:

1. Is there a /llms.txt file at the domain root? — a plain-text Markdown file that gives AI agents a curated reading list of the site. We checked by requesting https://${domain}/llms.txt and checking for a 2xx response with valid Markdown content.

2. Is there a /.well-known/mcp.json manifest? — the emerging convention for declaring which actions a site exposes to agents. Same check: a single HTTP request to the well-known path.

3. Are interactive forms annotated with WebMCP tool descriptions? — semantic descriptions of what each form does, so an AI agent can fill it out autonomously. Checked by inspecting the HTML for tool-description attributes on contact, search, and inquiry forms, and by looking at view-source for any registered tools.

4. Is content delivered via a queryable structured schema, or rendered as opaque HTML only? — checked by attempting to find a public API endpoint (GraphQL, JSON, or otherwise), by reading the rendered HTML for typed JSON-LD with property-specific schema, and by checking the document head for content-API headers.

5. Is the accessibility tree usable for autonomous interaction? — checked through Chrome devtools' accessibility panel for primary nav and form-control coverage, and by running Lighthouse 13.3 against the homepage and one listing detail page each.

The five sites

Anonymized so this lands as analysis rather than a callout, but representative of categories we work with every week:

Site A — a boutique developer with roughly a dozen active projects. Marketed on its design, its restraint, and its singularity. Website built on a popular page builder, hosted on a managed platform, rebuilt within the last 18 months.

Site B — a master developer with several hundred active listings across multiple sub-brands. Multi-page site with a search index, integrated CRM forms, and a custom internal listings interface. Older codebase, multiple agencies have touched it.

Site C — a high-end broker focused on prime resale. Around forty handpicked listings at any given time. Polished editorial design, slow scroll experience, no content beyond static project pages and a generic blog.

Site D — an off-plan agency aggregating launches from multiple developers. Volume site with hundreds of properties, comparison tooling, and aggressive lead-capture flow. Built on a template framework with heavy custom JavaScript on top.

Site E — a secondary-market specialist with a dedicated buyer-journey tool. Smallest of the five by listing count, but the most invested in editorial content (a content library, neighbourhood guides, a podcast feed).

All five rank in Google for relevant search terms today. All five run on substrates that pre-date the agentic shift. All five score zero against the items above.

What we found

/llms.txt: not present on any of the five domains. Each request returned a 404, which means Lighthouse 13.3 will treat that audit as Not Applicable rather than a failure — but the practical effect is the same. AI agents have no curated reading list and will pick URLs at random when they crawl.

/.well-known/mcp.json: not present on any of the five domains. Same 404 across the board. No site declares its actions to agents.

WebMCP tool descriptions on forms: zero across all five. Contact forms, search forms, viewing-request forms — none annotated. An AI agent crawling these sites can read the labels in plain English (so the accessibility tree is doing some work), but there is no programmatic description of what the form does, what kind of data it expects, what response it returns. We covered why that matters specifically in What is an agentic web platform? — the gap between a form an agent can READ and a form an agent can INVOKE.

Structured content via a queryable schema: zero across all five. Every site ships content as rendered HTML only — no public API, no JSON-LD beyond generic Organization or BreadcrumbList schemas, no headless CMS exposed. Site B has an internal listings API its frontend consumes, but it is not public. Site E ships JSON-LD Article schema on blog posts but nothing for properties. There is no path on any of these sites for an agent to query 'find me 3-bedroom apartments in Palm Jumeirah under 4 million AED' and get a structured response. This is the substrate question we discussed in What Is a Headless CMS? Why Sanity Powers the Best Modern Websites.

Agent-readable accessibility tree: mixed. Sites A and C score reasonably on basic accessibility (labelled controls, proper heading hierarchy, sensible reading order). Sites B and D have serious issues — heavy reliance on click-handlers attached to non-button elements, missing labels on filter controls, modals that trap focus incorrectly. Site E sits in the middle. Lighthouse 13.3's agent-readiness audits flagged accessibility issues on three of the five sites that would meaningfully degrade agent interaction even if the structural items above were addressed.

What this costs in the next 12 months

These are the three concrete losses we expect each site to absorb between now and mid-2027. None are theoretical. All trace to mechanisms that already exist or are documented to launch in the rollout schedules Google and Anthropic have published. We laid out the broader business case in AI Changed Your Business. You Just Haven't Noticed Yet, but property-specific:

Lost AI Mode citations. When a UHNWI buyer asks AI Mode about 'boutique developers Dubai Palm Jumeirah' or 'off-plan launches Business Bay 2026', the response will name specific brands. Being named requires the site to be readable in a way agents prefer. None of the five are.

Lost agent-mediated inquiries. Information Agents already monitor sites on a buyer's behalf — watching prices, new launches, availability changes. When the agent finds something its user might want, it will reach out on their behalf, which increasingly means submitting an inquiry through whichever form the agent can invoke. The five sites we audited present forms an agent can read but not invoke. The inquiries route around them.

Lost buyer leakage at the research stage. The research-to-contact gap on premium property sites was the subject of an earlier piece on why Dubai business websites aren't generating leads. The agentic shift widens that gap — buyers who would previously have submitted a form will instead let their agent gather the information and decide whether to surface the listing at all. If the agent can't get clean answers, the listing doesn't surface.

The fix path

For all five sites the fix is the same in shape, different in surface effort. The shape: serve a curated /llms.txt that agents can read, expose a /.well-known/mcp.json that declares the search, contact, and viewing-request actions, annotate the forms accordingly, and ensure the accessibility tree is clean enough for agent click reliability.

On a structured-content substrate, this is a route handler, a manifest file, and a small set of HTML attributes. Two to three days of focused engineering. The substrate is already shaped right; you are just exposing what's already there.

On a traditional CMS substrate, the same work is a parallel infrastructure problem. You hand-maintain /llms.txt (regenerating it every time content changes), you hand-write tool descriptions for each form (and update them every time the form changes), you hand-author structured JSON-LD that has to stay in sync with what the rendered HTML actually contains. Same outputs, every one of them now living in an additional system that has to stay synchronized. This is the maintenance cost that compounds the substrate question every quarter.

For property specifically the smart move is to start with the search form — it's the highest-traffic interaction surface and the easiest WebMCP tool to define cleanly. A description like:

{
"name": "search_properties",
"description": "Search the active property catalog by location, bedrooms, price range, and project status. Returns property name, location, price (in AED), bedroom count, completion date, availability status, and direct URL.",
"input_schema": {
"type": "object",
"properties": {
"location": { "type": "string" },
"bedrooms_min": { "type": "integer" },
"price_max_aed": { "type": "integer" },
"status": { "type": "string", "enum": ["off_plan", "ready"] }
}
}
}

Once that exists at /.well-known/mcp.json (or attached to the search form via declarative WebMCP attributes), an Information Agent monitoring property launches on a buyer's behalf can invoke it directly. The site becomes a function call. The form stops being the bottleneck.

What's next

We're publishing a longer piece this week on the full five-layer model — the architecture beneath an intelligent website, of which agent-readiness is one specific layer. After that, we open consultations on agent-readiness audits for Dubai property and hospitality businesses. Three slots in June, by application. Details at shapeshifters.dev/contact.

If you operate one of these sites and recognized your category, the substrate question is the only one that matters. Everything else is downstream of it.