How AI search reads a brief, and when to use quick search instead

The AI search box takes a sentence and turns it into a search. That sounds like magic and is mostly plumbing, so this note explains the plumbing. Knowing how it reads a brief makes it more useful, and makes it obvious when the quick search is the better tool.

Three layers

A brief passes through three layers before you see results.

The model reads the sentence. It fills a fixed set of fields: place, sale or rent, bedrooms, bathrooms, a price floor and ceiling, property type, a year-built window, and two lists of words. The first list is features a photo or listing could confirm: pool, ocean view, garage, Spanish style. The second is wishes nothing in the data can check: good schools, short commute, quiet street. Those become a note on the results page that says, honestly, "not filterable yet".

A set of rules checks the model's answer. Some things a model gets wrong often enough that we do not trust it alone. A street name in the place field is moved to a street filter, because the index treats a street as a single-property lookup and returns nothing. "United States" as a place means no place. "Under 600" on a home for sale means six hundred thousand. "No HOA" is an exclusion, not a feature to look for. A bare "Springfield" is ambiguous, so the box asks which one and offers the likely ones as buttons.

A plain reading if the model is unavailable. When the model cannot be reached, a simpler reader takes the sentence apart with rules: it finds the place, the bed count, the budget, the rent-or-buy intent, a few common features. You get a search with a small "basic reading" chip instead of an error.

The questions it asks back

The box will stop and ask when, and only when, searching would be pointless without an answer.

• No place at all: "Where should I look?"
• Several places: one button per place, so "Tahoe or Big Bear" becomes two searches you can pick between.
• An ambiguous city: which Springfield, which Portland.
• Not a home search: it says so and shows example briefs instead of searching the whole country.

Everything else it searches with what it has and tells you what it left out.

Refining a search you are already looking at

On a results page the same box becomes a refiner. It merges your follow-up into the search on screen instead of starting over. "Under 2M" changes the ceiling and keeps everything else. "Cheaper" lowers the ceiling by about a fifth. "No pool" removes the pool requirement; "I don't want a pool" excludes homes that mention one. "Try Malibu" moves the whole search. "Start over" does what it says.

If a refinement would empty the page, the box keeps your current results and offers ways to loosen: a higher ceiling, fewer bedrooms, any property type, rentals instead. Each is a full search you can run with one tap.

When quick search is the better tool

The AI box is for briefs: sentences with several constraints, some of which are judgement calls. If you know the city, the bed count and the ceiling, the quick search is faster and there is nothing for a model to interpret. Use the tab that matches the shape of what you know.

Where the AI box earns its keep is the style and feature words. "Mid-century with a view" is not a filter the index has. The box turns it into a year window plus feature words, and then a second pass looks at each candidate's photo, description and tags and ranks the strongest matches first, with a one-line reason on every card. That pass is slower and costs more, so it only runs when the brief has that kind of intent. A plain "3 bed under 600k" never triggers it.

One thing to remember

The box never guesses a place. If you leave the place out, it will ask. That is on purpose: a search of the wrong city, presented confidently, is worse than a question.