How AI will change property discovery

Property search has looked the same for twenty years. A map, a set of filters, a grid of results, a shortlist. The technology underneath has improved enormously. The interaction model has barely moved.

That is not because nobody has tried. It is because the filter-and-list model matches what a database can do, and until recently that was the constraint that mattered. It has never matched what a buyer is actually doing.

Filters describe inventory, not intent

Ask someone what they are looking for and you will not get a filter set. You will get something like: we need a third bedroom before the baby arrives, my partner cannot face another hour-long commute, and we would like to stay near her parents if we can.

Now watch that get entered into a portal. Three beds. A price ceiling. A radius drawn around a postcode. Everything that made the requirement specific — the trade-off between commute and proximity to family, the fact that the third bedroom could be a decent loft conversion, the timing pressure — is discarded, because there is no field for it.

The portal then returns four hundred results, and the buyer spends six weeks manually reintroducing the context the form threw away.

This is the actual failure. Not ranking quality, not photo quality, not data coverage. The search interface cannot represent the requirement.

What changes

Language models are good at precisely the thing the filter form is bad at: holding a messy, multi-dimensional, partially-contradictory requirement and reasoning about trade-offs within it.

That produces four shifts.

Search stops being a form and becomes a conversation

The requirement above can be stated as it was actually thought. The system can then ask the question a good agent asks — how much commute would you trade for being ten minutes from her parents? — and use the answer. Preference elicitation becomes part of the search rather than something the buyer has to complete before searching.

Listings get read, not just indexed

Most of what determines whether a property fits is in the description, the floor plan and the photographs, none of which are currently searchable in any meaningful way. “South-facing garden”, “needs modernising”, “loft already boarded with a fixed staircase” — this is the material buyers actually decide on, and it sits in unstructured text and images that the filter model cannot reach.

Vision and language models can extract it. The searchable surface of a listing gets several times larger without a single agent filling in another field.

The neighbourhood becomes part of the property

Buyers are not purchasing a building, they are purchasing a daily life. Commute times at the hours they actually travel, school catchments that will apply when their child is old enough, whether the high street is improving or hollowing out. This information exists, scattered across a dozen sources. Assembling it is exactly the kind of tedious synthesis that models are now good at.

Discovery becomes continuous

Right now, search is an activity you perform. The better model is a standing brief that works while you do not: the requirement is held, the market is monitored, and you are contacted when something genuinely matches — including when a property that has been listed for months drops into range.

What this does to the industry

Here is the part that should concern anyone running a portal.

The current business model rests on being the place where inventory is aggregated. That moat holds while search is a browsing activity, because browsing requires somewhere to browse.

If discovery becomes a conversation with an assistant, the assistant becomes the interface — and it does not care whose database it queries. The aggregation advantage does not disappear, but it stops being customer-facing, and a supplier of data to somebody else’s assistant captures far less value than the destination did.

Portals that treat AI as a feature to add to the search page are solving for the wrong thing. The question is not how to make filters smarter. It is whether the portal is still where discovery happens at all.

What to actually do

For anyone building in this space, the sequencing matters more than the ambition:

1. Start with listing enrichment. Extracting structured attributes from descriptions, floor plans and images improves the existing product immediately and is a prerequisite for everything else. Do this first regardless of what you believe about the rest.
2. Instrument intent. You almost certainly do not capture why a buyer rejected a property. That signal is the most valuable data in the business and most portals discard it.
3. Treat conversational search as a distinct surface, not a chat widget bolted onto the results page. The two have different interaction models and cannibalising the second with a bad version of the first helps nobody.
4. Decide, at board level, whether you intend to be the assistant or to supply one. Both are viable strategies. Drifting into the second while believing you are pursuing the first is not.

The buyers were always doing something more sophisticated than filtering a list. The tooling has finally caught up with them.