The Rise of Natural Language Property Search

Published on offmarket.now | July 2025

For twenty years, searching for property in Australia has meant the same thing: select a suburb, set a price range, choose bedrooms and bathrooms, pick a property type, click search. The interface has barely changed since Domain and realestate.com.au launched their first search pages.

In the last twelve months, that paradigm has started to break.

What natural language search means

Natural language search lets you type (or speak) what you're actually looking for, in plain English:

• "A quiet house near the beach with a big backyard, under $1.2 million"
• "Investment apartment in a walkable neighbourhood with good public transport"
• "Three-bedroom terrace in the Inner West with parking, renovated kitchen"

The system interprets your intent, maps it to property attributes, and returns relevant results — even when your query doesn't map neatly to database fields.

This isn't keyword search. It's semantic understanding. The system knows that "quiet street" implies low traffic, that "walkable" means proximity to shops and cafes, that "renovated kitchen" can be inferred from listing photos and descriptions.

Who's building it

REA Group led the charge in Australia, rolling out natural language search to all realestate.com.au users in late 2025 after a year-long pilot. Their February 2026 ChatGPT app took it further, embedding property search directly into a conversational AI platform.

In the US, Homes.com launched AI Smart Search in October 2025, powered by Microsoft Azure OpenAI. Realtor.com followed with its own natural language tool the same month. Roof AI has been building natural language search specifically for real estate verticals.

CoStar Group, which completed its acquisition of Domain in August 2025, has described its Homes AI as "the most sophisticated vertical AI application in real estate." When (not if) this technology flows into Domain's Australian platform, the country's two largest property portals will both offer conversational search.

Why filters fail

Traditional filter-based search has a fundamental limitation: it can only match against structured data fields. If a property has three bedrooms plus a study, a search for "four bedrooms" won't find it — even though the study might work perfectly as a fourth bedroom.

Filters also can't capture:

• Lifestyle preferences — "family-friendly," "young professional area," "low maintenance"
• Qualitative features — "lots of natural light," "character home," "modern finishes"
• Relative criteria — "bigger than my current place," "closer to work than where I am now"
• Composite requirements — "good schools AND walking distance to shops AND under 15 minutes to the CBD"

Natural language search handles all of these by understanding intent rather than matching fields.

The off-market angle

Off-market properties are particularly poorly served by filters. They often have:

• Incomplete data — no professional marketing campaign means fewer structured fields are populated
• Non-standard features — unusual layouts, mixed-use potential, development opportunity
• Contextual value — "this property is valuable because it's next to a planned metro station" is intent that filters can't express

When off-market listings are surfaced through aggregation platforms, they frequently lack the polished, structured data that portal listings have. Natural language search can work with fuzzier inputs — a partial address, a description, a set of characteristics — to match buyers with relevant properties.

For platforms that aggregate off-market listings from agency websites, the combination of data aggregation + natural language search creates a discovery experience that filter-based portals can't replicate for the off-market segment.

The technical reality

Building natural language search for property is not trivial, but it's more accessible than it was even a year ago:

• Embedding models convert property descriptions and buyer queries into numerical vectors that can be compared for similarity
• Large language models interpret ambiguous queries and extract structured intent
• Retrieval-augmented generation (RAG) combines a property database with an LLM to answer specific questions about listings
• Multimodal models can factor in listing photos — identifying renovated kitchens, pools, and views from images rather than relying on text descriptions

The key constraint isn't the AI — it's the data. You need a comprehensive, geocoded, regularly updated listing database for the search to work against. This is why aggregation platforms (which maintain their own database by scraping multiple sources) are well-positioned to offer natural language search.

What this means for your property search

• Start using conversational search tools — REA's ChatGPT app is free. Try expressing what you actually want, not what filters allow you to say
• Don't abandon traditional search entirely — filters are still faster for simple, well-defined queries ("three-bed house in Bondi under $2M")
• Set up alerts for your target postcodes — conversational search is great for exploration, but alerts catch properties the moment they appear
• Express your real criteria — don't self-censor to fit filter categories. If you want "a fixer-upper on a big block near good schools," say that

The property search interface hasn't meaningfully changed in two decades. It's changing now, and quickly. Buyers who adapt will find properties that others miss — especially in the off-market segment where traditional search has always fallen short.

---

Browse off-market listings or explore by suburb on offmarket.now.