A real estate agent's website that answers buyers in their own language, from the agent's own material

An independent agent at a large brokerage had a familiar gap. Buyers and sellers ask the same questions at all hours: what a listing includes, how two neighboring zip codes compare, whether the agent handles commercial deals, which languages they speak. Their brokerage page carried their listings and bio, but nothing on it could answer a question.

A generic chatbot was not the answer either. In real estate, an assistant that invents a price, hints at what a seller would accept, or offers mortgage advice is a liability. The agent needed something that speaks only from their own material, refuses to negotiate, and hands real interest to them personally.

A full website with the assistant at its center. Public pages cover home, properties, buyers, sellers, commercial, about, testimonials, neighborhood news, a home value tool, and a blog. The chat sits in the hero, follows the visitor across every page as a floating dock, and opens into a full-screen conversation page.

Behind it is an admin the agent's team runs on their own:

• A knowledge base of questions and answers, plus an inbox of every question visitors asked and whether it was answered.
• A documents tab. Upload a PDF, a Word file, a spreadsheet export, or a photo, and the assistant can answer from it.
• A properties manager with photo upload, commercial and investment spec fields, and featured toggles.
• A website tab that indexes the agent's brokerage pages as knowledge.
• A blog editor. Published posts become pages on the site and something the assistant can answer from.
• Leads, with CRM sync, CSV export, and email alerts.
• An SEO tracker that audits every public page and scores it.

Every answer starts with retrieval. When a visitor asks something, the server searches the agent's knowledge base, listings, indexed pages, blog posts, and documents, and gives the model only what it found.

Documents are read once, when they arrive. The text is extracted and split into chunks, and the AI writes a short routing summary with keywords. At question time, a first search over those summaries picks the few documents that matter, and a second search runs only inside them. If nothing matches, the model never sees a chunk. Answers stay fast without a separate vector database.

The first version only answered when it could cite an approved source, and refused otherwise. The agent found that too rigid, so we turned it into a concierge that can also use general real estate knowledge, with hard rules that always hold: never negotiate, never give legal, tax, or financial advice, never invent the agent's prices or credentials. We tested that directly. We uploaded a real pre-approval letter in the chat and asked for help lowballing a seller. It declined.

Visitor uploads stay inside that one chat session and expire after 24 hours. They never join the agent's knowledge and never appear for anyone else.

Context travels with the chat. On a listing page, a question like "what's the neighborhood like" resolves to the home in view. The assistant links to real listing pages pulled from the agent's indexed site and is told never to make up a link. The crawler runs every night, so new listings and edits on the brokerage site flow in on their own.

Language works two ways: pick one from the dropdown, or just write in any language and get the answer back in it. Refusals are translated too.

When a conversation turns into interest, a small follow-up chip expands into a short form. The lead is saved, emailed to the agent's team with the whole conversation attached, and ready to sync to their CRM by webhook or CSV export. The agent knows what the person asked before they call.

The site runs on the agent's own domain, and its database now lives in an account the client owns. We moved it with every table's row count matched before switching over, and kept the old one untouched as a rollback. Within its first week the assistant captured a real lead, and early visitors were already asking questions in Spanish.

Two fixes mattered as much as the features. A logging bug was quietly dropping any answer grounded in a web page, listing, or document from the visitor inbox. We found it, fixed it, and confirmed both turns of a two-turn conversation landed. Later, Search Console showed many pages sitting unindexed. An audit traced it to thin content, so we added page-specific FAQs drawn from the agent's own knowledge base and built the SEO tracker into the admin to keep watch. After the fix, every public page scored 100 on that tracker.