Most "AI search" is a chatbot bolted to the corner of a page. For Rye & Beyond Cottages, a holiday-lettings brand on the East Sussex coast, we built the opposite: an AI concierge that drives the actual search interface. It filters the catalogue, checks live availability, plans the trip, and writes the results straight onto the page. And the conversational layer, the part everyone assumes is hard, we barely wrote. That's the bit Chat Thing does.
Here's the twist up front, because it's the whole point of this post: the AI was the easy part. Stick with me and I'll show you why.
The widget on load. Chat on the left, the full catalogue on the right. One surface, two ways in.
Holiday-lettings search is three hard problems in a trench coat
You've got a catalogue people want to browse and filter. You've got fuzzy, human intent ("somewhere dog-friendly near the beach with a woodburner for a long weekend") that keyword filters can't read. And you've got availability, which changes by the day and makes half of a beautiful results page irrelevant the moment someone picks dates.
Most sites pick one and fake the rest. Rye & Beyond wanted all three to work together, and they wanted it to feel like talking to someone who knows the cottages rather than filling in a form. So we built a single embeddable surface where a guest can browse and filter the cottages, search them by meaning, see live availability and real stay prices for their dates, and talk to a concierge that operates the whole thing for them.
The concierge drives the real UI, it doesn't sit beside it
This is the bit that makes it feel different. The assistant doesn't answer in a side panel and leave you to do the clicking. It operates the actual interface.
We gave the concierge a set of tools, registered as Chat Thing power-ups, and each one maps to something a person can do on the page: record a preference, set the stay dates, run a search, write the results into the grid, open a cottage, build a shortlist, compare two side by side, plot what's nearby, plan an itinerary, email it. Ask it for "a dog-friendly cottage in Rye that sleeps four" and it records the preferences, checks live dates, runs the search, and sets the results. Every step is a visible tool call in the thread, so using it reads like watching someone competent use the site.
Ask in plain English and watch the grid answer. The assistant records preferences, checks live dates, sets the results. Each step is a visible tool call.
Because the tools are the same actions a human has, anything the guest does by hand and anything the assistant does end up in one shared state. Filter by hand, the chips update. Ask the concierge, the same chips update. There's no second copy of the truth, and that single decision is what stops the whole thing feeling like a bolt-on.
Search by meaning, not keywords
The "semantic search" on most sites is a keyword match with better marketing. Ours is real. Every cottage is indexed for vector search, so when someone searches "romantic bolthole with a woodburner", the natural-language query is ranked by meaning and mapped onto the live catalogue. Not a filter pretending to understand. An actual ranker doing the understanding, with the editable catalogue still in charge of the facts.
Availability, done honestly
This is the part that quietly makes or breaks a lettings site. We overlay a live availability check on the catalogue: give it dates, a party size and whether a dog's coming, and it returns exactly which cottages are bookable and the real total for that stay.
Then we show everything. Available cottages get a badge and a real stay price. Booked ones stay on the page, greyed and clearly marked, so a guest never hits a dead end or wonders why their favourite vanished. A "show available only" toggle clears the rest when they're ready to commit.
Every cottage tagged against real dates. Available ones show a total stay price. Booked ones stay visible but greyed, so nobody hits a dead end.
We were careful about one thing, and it matters more than any feature: the assistant never guesses. If no dates are set, it won't call anything "available". It says so, offers to check, and only claims availability once the live lookup has actually run. An AI concierge that confidently books you a cottage that's already taken is worse than no concierge at all. Open any cottage's calendar and you get its own live view, booked days in red, free days in green, pulled per cottage so it matches what's actually bookable.
Every cottage carries its own live calendar, fetched per cottage so it always matches reality.
Turning a search into a trip
Once a guest is leaning towards a cottage, the concierge can plan the holiday around it. It pulls local events happening during the stay and the nearest beaches, pubs, restaurants and attractions, plots them on the map with distances from the front door, and writes a day-by-day itinerary you can actually scan.
The assistant builds the itinerary from real events and nearby spots, then plots it as a timeline you can read day by day.
And the guest can have the whole thing in their inbox: a branded itinerary email with the cottage, the dates, what's on, the day-by-day plan, nearby highlights with distances, and a button straight to booking.
One tap and the plan is in their inbox. Branded, dated, with a booking link, every stop with a distance from the door.
The small things that make it feel finished
The headline features get you in the door. The small ones make it feel like a real product rather than a demo. Ask the concierge to compare a few cottages and it flips to a side-by-side with honest pros and cons for each, drawn from the real data: which sleeps the most, which is cheapest per night, which has parking, which is closest to the things you asked to be near.
Ask to compare and the assistant lays them out side by side, with real pros and cons per cottage, not marketing fluff.
Cards carry quiet "good to know" notes scanned from the house rules, steep stairs, an unfenced garden, a pond near where the kids and the dog will be. On mobile the chat becomes a slide-up sheet over a results-first screen, so the assistant is always a tap away without getting in the way. None of it shouts. All of it adds up to a search you can trust.
On mobile the chat is a slide-up sheet. Results first, assistant a tap away.
The part that surprises people: the AI was the easy bit
Building an assistant that drives a real app sounds like the hard part. It isn't, because the conversational layer runs on Chat Thing.
Chat Thing handles everything that's genuinely fiddly: the chat interface, the conversation state, the model, streaming, and the secure handshake into the page. What we bring is small. We register each of the app's actions as a power-up, a function plus a short description of its inputs, write a system prompt that gives the concierge its persona and house rules, and push the current page state as context. Load the SDK, register a handful of tools, override the prompt. That's the integration.
So the effort goes where it should: into the product. The search, the availability logic, the itinerary, the map, the details. The AI is a thin, well-understood layer on top, not a research project. That's also why the pattern moves to the next business cleanly. Swap the tools for that company's actions, rewrite the prompt, and the same agentic front end comes along for the ride.
Want this for your catalogue?
If you run a business with a lot of inventory and customers who don't quite know how to filter for what they want, an estate agency, a travel brand, a marketplace, a big shop, this pattern fits. An AI concierge that drives a real, filterable, bookable surface, with honest live data underneath, and the conversational layer handled for you by Chat Thing.
We help companies design and build exactly this. If you'd like one for your own catalogue, get in touch and we'll talk through what it would take.
Not a chatbot in the corner. A search experience that happens to be agentic all the way down. 🐰
Built on Chat Thing (the conversational layer), by the team at Pixelhop.