Describe the home. The filters fill in — and show their work.
Type or speak in any language. Location, sale or rent, property type, price, bedrooms and bathrooms become real filters. Preferences only change the order and never remove homes. An editable summary shows exactly how the request was interpreted.
Text or voice
Two ways to describe the same search
A place stays a place; a description becomes a full request
“Brickell” or “33139” uses the instant place lookup. Three or more words, any digit or currency symbol, or a continuous CJK or Thai string is treated as a descriptive request. It first reaches the website’s own route and is then interpreted by the WebGrowly hub.
The standard place lookup remains in the same panel for buyers who already know the address or neighborhood.
A voice clip with automatic language detection
The browser records for up to 60 seconds with MediaRecorder and uploads the clip. An AI transcription model detects the language; the browser’s Web Speech API is not used. That keeps the same recording flow in Chrome, Safari and Firefox without asking the visitor to choose a language first.
Accepted formats: webm, ogg, mp4, m4a, mp3 and wav. Typed and spoken versions share the same request cache.
Request to results
How a sentence becomes filters and ranking preferences
- 01
The buyer opens the language search
One control on the search page opens a centered panel on desktop or a full-screen sheet on mobile, with examples and a microphone button.
- 02
The request separates filters from preferences
Location, ZIP, sale or rent, property type, price, bedrooms and bathrooms become hard criteria. Up to 8 preferences can come from a closed list of 52 concepts. A fixed multilingual table checks the model’s result again against word-root patterns in English, Spanish, Russian, French and Portuguese before anything is labeled “couldn’t use.”
- 03
The standard search receives the hard criteria
The same filter state, URL, result list, map and Market Pulse used by the classic search stay in place. Preferences only reorder those results; they never remove a property.
- 04
Evidence is labeled by strength
A positive structured MLS field is “confirmed.” A match found only in the broker’s description appears as a quoted sentence, not a checkmark. A negative field is “contradicted.” A blank field is “unknown” and is left out of that property’s denominator rather than scored as a failure.
- 05
The buyer receives an editable interpretation
The summary separates applied filters, ranking-only preferences and anything that could not be used. Every tag can be removed. Removing a filter runs the search again; removing a preference changes the highlights and order, not the result count.
The request and the interpretation shown together
The returned tags distinguish real filters from ranking preferences and identify anything the system could not apply.
Exact rules
The preference list, limits and language handling
- Preference list
- A closed set of 52 concepts. Anything outside the set cannot be scored, labeled or repeated back as a preference.
- Named examples
- furnished, washer/dryer in unit, corner unit, high floor, penthouse, renovated, pool, gym in building, elevator, doorman, gated, balcony, boat dock, ocean view, bay view, oceanfront, impact windows, pets, no HOA, near beach, near transit, walkable.
- Limit per request
- A single request can carry no more than 8 preferences.
- Effect on results
- Order only. Adding or removing a preference does not change the number of properties.
- Preference-only requests
- “Furnished, gym nearby” is a valid whole-market search that changes the order instead of returning an error.
- "Near the beach"
- Measured geometrically against the embedded shoreline from the listing's own coordinates, not guessed from marketing text.
- Input languages
- Requests can be entered in any language. City names and listing terms are converted to the English values expected by the MLS query.
- Request routing
- Three or more words, any digit or currency symbol, or continuous CJK or Thai text uses the language interpreter. Shorter input stays with instant place lookup.
- Voice clip
- Recorded by MediaRecorder in the browser and stopped at 60 seconds; webm, ogg, mp4, m4a, mp3 or wav.
- Bedroom counts
- Accepted only when the sentence actually contains a bedroom word in one of the supported languages, so the bare number in "hasta 3000 al mes" is read as a price, not a bedroom count.
- Named neighborhoods
- Resolved to a real boundary polygon before searching, because in the MLS feed the City field for Brickell, Edgewater or South Beach says "Miami".
- Shared cache
- Voice and text versions use the same request cache.
- No AI key configured
- The control falls back to the standard place search instead of presenting an AI action that cannot work.
What this interpretation never does
- Protected-class language is removed before it can filter, rank or be repeated back, in all five website languages.
- School quality is acknowledged but never ranked, and nothing in the search can select for or against people, neighborhoods or safety.
- An accessibility need — step-free, wheelchair, elevator, ramp — is surfaced with a note telling the visitor to raise it with the agent. It is neither silently discarded nor used to rank.
- Interpretation and transcription are automated, not human translation, so no arbitrary request is guaranteed to be understood.
- Wishes are not filters. Nothing is removed from the result set because of one.
- The AI does not find homes on its own, contact anyone, or book anything, and nothing here is a claim about search rankings, traffic or lead volume.
Place in the product
Another way into the same property search
Bring a search request in your own words
Type or speak it during the demo and inspect the filters and preferences the system returns.
Book a Demo