A framework for discovery
Some goods can't be ranked by what sold before. Resonance ranks them on what they actually are — and on what the person in front of you actually asked for.
See it in action at CielStay →The problem
A long-term rental. A classified ad. An artwork. A custom camper van. A one-off holiday house. Each is close to unique, bought once, and judged on what it actually is. People like you also chose this has nothing to work with — there is no second unit of the same thing, and yesterday's winners say nothing about today's fit.
Collaborative filtering isn't broken. It's precise in its own regime: interchangeable items, bought repeatedly, by people whose history predicts their next move. Applied to goods judged at face value, it degenerates. Items with history outrank items without. The newest and most unusual listing comes structurally last. Cold start isn't a launch problem here — it's the permanent condition of the inventory.
The framework
Resonance is built on a simple observation: consumers know what they want — they just have no good way to say it. A guest doesn't want "3-bedroom house near beach." They want something ineffable: the feeling of waking up to fog, somewhere quiet but not isolated, with room for a dog and a long porch. They'll describe it if you let them.
On the producer side, every product contains rich, latent structure — qualities, textures, experiences — that never make it into a search index. A vacation rental isn't just bedrooms and bathrooms. It's the light in the morning, the distance from a trailhead, the vibe of the neighborhood.
LLMs are the first technology capable of holding both sides of this simultaneously — parsing ephemeral consumer signals and comparing them against structured producer reality — at scale and in real time.
How matching works
Resonance matching is not a single comparison — it is four complementary signals brought into alignment. Two from the consumer. Two from the producer.
Legacy systems match query against description — one signal on each side. Resonance matches all four simultaneously. The result is not just relevance, but genuine fit: the right guest for the right property, at the right moment.
Resonance was inspired by the style and works of Edith Wharton — the novelist who understood, better than most, that the distance between what people say they want and what they actually need is where all the interesting work happens.
Core principles
01
A query is not a bag of words. It is an expression of latent desire. Resonance systems parse meaning, not tokens — which means they work even when users don't know the right vocabulary.
02
Vagueness isn't a virtue. Resonance extracts specific, structured attributes from natural language — then matches them against specific, structured product reality. The goal is a tight, explainable fit — not a probabilistic shrug.
03
Consumer signals are richest before they're cleaned. A direct quote — "I want somewhere my kids can be loud without me worrying" — contains more signal than any checkbox. Resonance preserves and uses raw, unstructured intent.
04
What you want today is not what you wanted last month. Resonance does not build a profile over time. It takes what you volunteer right now, uses it completely, and discards it. Every session is a fresh conversation.
Why it matters
I
Not cold start eliminated — cold start assumed, always. Nothing is ever promoted for having been chosen before. Rank comes from what the item is and what the person asked for, and from nothing else. No profiles, no history, no winners' circle. Every product is discoverable on its first day and on its thousandth.
interaction-free by commitment, not by necessity
II
Traditional recommenders are surveillance machines. They require persistent behavioral tracking, cross-session identity, and accumulated history to function. Resonance is ephemeral by design — no profile is built, no history is retained. The system is fully explainable: every match comes with a stated reason, not a black-box score.
III
Legacy search can't accept natural language from an AI agent. Resonance was designed for exactly this handoff — structured intent extraction from unstructured language is the core primitive.
But the opportunity goes deeper. Agentic AI doesn't just relay your query — it understands where you're coming from. It can construct a rich, privacy-friendly persona from your expressed needs, tastes, and desires. It knows why a midcentury modern home with Hollywood Regency decor is true to character in Palm Springs. It recognizes when a treehouse is a work of craftsmanship rather than an opportunistic novelty. That contextual intelligence, paired with Resonance's structured producer data, produces matches that no keyword search — and no human concierge at scale — could achieve.
For the technically inclined —
Retrieval runs on bi-encoders — a guest's intent becomes a vector, a property's character becomes a vector, and matching is the geometry between them. That part is unremarkable, and we say so deliberately: the encoder is not the differentiator. Two generations of fine-tuning it on our own corpus changed end-to-end results by nothing measurable. We run a general-purpose model and spend our effort elsewhere.
What differentiates Resonance is what gets encoded. Standard retrieval indexes what a document says. We don't index what a seller writes about themselves at all — rank that follows a seller's own copy rewards copywriting. Instead every item is re-described through one model, applying the same mildly scrutinising eye to all of it, synthesised from prose, visual analysis and market context. Ranking happens on that. It is not a marketing description. It is a judgment.
That normalisation layer is the load-bearing component, and we can put a number on it: changing which model re-describes a property moves its median rank by a factor of nine. Changing the embedding model moves it by nothing. Most semantic-search work optimises the second number.
Ordering is then settled by a cross-encoder that reads each candidate against the query directly, rather than by the vector geometry alone.
What we trust
A system that refuses behavioural data has to be explicit about what it does rely on. We extend trust in exactly three places:
I
Sellers, and vision models reading their photographs, know their own goods better than any behavioural proxy does.
II
Directly, or through the AI agents now acting on their behalf. Volunteered in the moment, not inferred from a file.
III
Whether this genuinely answers that — the one job neither party can do for itself.
And we verify exactly two things: that the listing exists on a platform we can inspect, and that whoever claims it owns it. Nothing else is taken on faith, and nothing else is required.
The honest part
Every architecture trades something. These are ours, and we would rather state them than be found out holding them.
Remove popularity bias and description quality takes its place. Measured on our own catalogue: at equal quality scores, listings carrying our better prose get roughly twice the exposure of those still on legacy text. The answer is uniformity — one describer for the whole corpus — which is why we rewrite rather than accumulate.
We hold no behavioural check on an inflated claim. The photographs are our only independent evidence, so cross-modal agreement between what a listing says and what its images show is the verification channel we are building.
A system that ignores past outcomes cannot learn from them. We can be consistently wrong and never notice. That is the price of refusing to pick winners, and we would rather pay it than rank the long tail last.
First deployment
The first production application of the Resonance framework. A semantic search layer over the world's most comprehensive independent vacation rental database — prioritizing unique, character-rich properties over mass-market commodity listings. Every property carries a written character study and sits in one of 5,666 local markets, with 2.5 million photographs behind them.
What comes next
Vacation rentals are the first domain. Any market where consumer desire is complex and producer data is rich — real estate, travel, retail, talent — is a candidate. The framework generalizes. The moat is the data and the methodology.