n

From the lab — part of VR AI Labs, where Virtual Results designs AI for real estate websites in public. A short read on what we’re prototyping, why, and what’s hard. Skim it, then scroll down and try the idea.

What if your clients could ask your market report a question?

Not a chatbot in a bubble in the corner of the screen. Not a generic AI assistant who doesn’t know whether they’re asking about Willow Glen or Westchester. A conversational layer anchored to the specific article they’re reading — one that pulls from the article text, from live MLS data, and from the reader’s browsing history to give answers that are actually useful.

This is what we’re calling Chat With This Article. It’s our most ambitious feature in the VR AI Labs series, and probably the one with the highest ceiling for both conversion and genuine client value.

The Problem With Every Other Chatbot

Standard real estate chatbots have a fundamental design flaw: they’re disconnected from the context the client is already in. Someone reading a detailed market analysis of Willow Glen doesn’t want to start a blank conversation with an AI that doesn’t know what page they’re on. They want to go deeper on this content.

Perplexity figured this out for web search. NotebookLM figured it out for documents. We’re building the real estate version: an AI layer that reads the article the same way your client does, knows which listings are relevant to the neighborhood being discussed, and cites its sources with inline references the client can verify.

How It Works

The experience has four distinct states, each designed to feel native to the reading flow:

  • Collapsed launcher: A slim strip at the bottom of the article section. “Chat with this article about Willow Glen →”. It’s visible but never intrusive — a gentle invitation, not a demand for attention.
  • Open panel: The agent’s presence is established immediately (Sarah Romero, in this example), with four suggested questions that match the article’s actual content. A context echo line below the input tells the client exactly what the AI has access to.
  • Mid-conversation: Inline superscript citations let the client verify any claim against the original article text. Hover a citation and you see the exact passage the AI drew from. This is crucial for MLS accuracy and client trust.
  • Property cards in answers: When the AI surfaces a listing, it shows as a proper card — address, price, beds/baths, with a “Show on map” action. Not a link. Not a reference. A card that invites a next action.

Interactive Demo

Click through the states below to see the full experience. Each tab shows a different moment in the conversation flow:

What We’re Still Figuring Out

Building this right requires solving three hard problems simultaneously:

  • MLS accuracy: The AI must only surface verified listing data. No hallucinated prices. No outdated inventory. This means real-time Typesense integration with strict grounding.
  • Context window management: Long articles can exceed practical context limits. We’re testing chunking strategies to ensure the AI can always find the relevant passage, not just the most recent one.
  • Agent attribution: When a client books a showing from a conversation that started in an article, how do we attribute that to the agent who wrote the article? This is a CRM integration problem we’re actively working through.

We want your perspective. Have you tried a similar feature? What questions do your clients actually ask when they’re reading market reports? What would make you more confident deploying AI that cites sources on your behalf? Share your thoughts in the comments — your real-world experience is exactly what helps us build the right thing.

This feature is currently in research and prototyping phase. The interactive demo above shows the target UX; implementation timeline depends on adoption signal from the VR agent community.

n

Lab notes: how we’re prototyping real estate chatbots that actually know the page

n

Consider this me thinking out loud from the workbench. The goal we set was narrow on purpose: not “build a smarter assistant,” but “let a reader interrogate this market report and get answers grounded in the article and live MLS data — with citations they can check.” That framing changes everything, because it turns a generic chatbot into a reading tool.

n

The honest hard part is grounding. The moment an AI talks about prices or inventory, a confident-but-wrong answer is worse than no answer at all — it erodes the trust the agent spent years building. So our prototype refuses to free-associate: every claim has to trace back to the article text or verified listing data, or it doesn’t get said. That constraint is slower to build than a chatty bubble. We think it’s the only version worth shipping.

n

What the evidence actually says about AI for real estate websites

n

We’re weighing this, not cheerleading it. The pattern is proven in publishing — and the failure mode is just as well documented.

n
  • Harvard Business Review’s generative-AI work — HBR’s own “Ask AI” reading feature shows subscribers will adopt and return to chat-with-the-content when it’s done well.
  • Google NotebookLM — the source-grounded model we’re emulating: answers stay tied to your documents, with inline citations for transparency and trust.
  • Stanford HAI on hallucination — the skeptic’s case: ungrounded models get specific facts wrong at alarming rates. This is exactly the risk our MLS-grounding constraint exists to kill.

See it live: try this and every VR AI Labs prototype in the Interactive Demo Showcase — live, clickable, on phone or desktop.

The VR AI Labs Series

A field guide to making AI a first-class citizen of the real-estate website — not a chatbot bolted into the corner. Explore the full series: