Replacing filter-based apartment search with a conversation

TL;DR — I replaced a 20-filter apartment search with a guided conversation — grounded in behavioral science and AI trust research — so renters could move from vague intent to confident shortlist.

Context

A filter panel assumes the user already knows what they want. Most renters don’t.

This project was never really about designing a UI. It was about integrating LLM behavior, RAG-grounded outputs, and behavioral psychology into a single experience that could earn the trust of Americans making one of the most financially and emotionally loaded decisions of their lives — finding a home. The interface was the last layer, not the first problem. And underneath it all was a second integration challenge just as hard: making a high-quality conversational experience feel seamless on top of rich, structured real estate data — so that the intelligence Markerr had built for institutions could finally speak to a renter in plain language.

Markerr had the data engine — rent forecasts, market comps, neighborhood signals already powering institutional investors. The next challenge was putting that same intelligence in the hands of renters across Markerr’s property management partners: hundreds of thousands of units, and a search experience that expected certainty from people who had none.

Ask Markerr became the answer — a conversational AI layer sitting on top of Markerr’s data engine, surfacing the right properties through dialogue rather than filter panels.

Role

I was the solo Product Designer on this engagement — research synthesis, information architecture, interaction design, and annotated handoff specifications across iOS mobile and desktop web, over approximately 20 weeks.

I worked in close collaboration with Markerr’s product and data teams and their property management partners’ digital experience teams.

The constraints were real from day one:

• Trust had to be earned quickly — users were handing personal context to an AI they had never met.
• Dealbreakers (price, pets, accessibility) could not be buried behind extra taps.
• The flow had to feel guided without removing the user’s sense of control.

Discovery

My research ran two parallel tracks from day one.

Track 1: How people actually search for housing in America. What I found disrupted my initial assumptions. Fifty-nine percent of renters are browsing with no firm move-in date — they are explorers, not decisive buyers. Their search is inherently social: couples have documented disagreements about budget (24%), space (22%), and duration (20%). Families spend weeks researching school districts before booking a single tour. First-time renters bring their parents. The decision is rarely made by one person, alone, in one session.

What struck me most was the consistency of the first question renters ask, across every segment: “What’s included in the rent?” Not neighborhood. Not square footage. Financial transparency before anything else.

Track 2: How Americans behave when interacting with LLMs. Over 52% of U.S. adults now use conversational AI — but 40.7% remain skeptical of what it tells them. Thirty-two percent rarely feel understood by chatbots. The root cause is a mental model mismatch: people bring human conversational expectations and the AI disappoints them. The most common failure modes aren’t bad answers — they’re poor error recovery, no visible reasoning, and no feedback that the system actually got what you meant.

Seen through a behavioral design lens, both problems share the same root failure: systems built for users who already know what they want. The filter panel assumes a certain user. The chatbot assumes a certain user. Most renters are not that user.

Reframing

After research, the brief quietly shifted. The original framing was: make apartment search feel more modern. The reframing was: design for the uncertain renter, not the certain one.

A feature-led redesign would have shipped faster and changed almost nothing — users would have brought the same uncertain intent to a more modern interface.

This changed everything downstream. It meant the interface couldn’t front-load choices. It meant the AI had to earn trust before showing results, not after. It meant dealbreakers had to be on the surface, not behind a tap.

The HMW that guided the design: How might we help renters who start with a feeling — not a spec sheet — reach a confident shortlist without becoming their own filter engineers?

Design Decisions

The flow maps a complete renter journey across eight screens. Each screen was designed around one behavioral principle.

Testing

The hardest micro-decision in testing was the conditional CTA timing: when exactly should the result count badge appear?

Early surfacing — before the echo — caused users to jump to results before the AI had enough signal. Late surfacing felt like the AI was hiding something. The echo-confirmation trigger was the resolution: badge appears only after the user sees their constraints reflected back.

One assumption testing broke that I hadn’t expected: I thought progressive disclosure of the result count — appearing gradually as the user answered questions — would feel rewarding. In testing it felt manipulative. Users wanted to know if results existed before committing to the full conversation. The echo-confirmation trigger was a better answer: show the count after I’ve demonstrated I understood them, not before.

One thing I would have instrumented earlier: drop-off data by question position in the chat screen. The one-question-at-a-time rule is evidence-backed, but I’d want to know which question in the sequence causes hesitation before considering any modifications in production.

Final Design

The three decisions that define the experience:

1. Echo before resolve — the AI restates intent before acting, the primary trust mechanism in the entire flow.
2. Dealbreakers on the surface — price and pet policy visible at the card level, no taps required.
3. Map as output, not filter — spatial reasoning is the dominant mental model; the map earns its position as a primary results view.

Impact

The outcomes were directional and benchmark-grounded. Markerr did not yet have production data at handoff — the numbers below are from the behavioral research that underpinned every design decision:

Metric	Benchmark source	Signal
Listings viewed	Redfin, Nov 2025	2× more with conversational vs. filter search
Tour request likelihood	Redfin, Nov 2025	+47% with conversational search
AI trust (transparency effect)	Francis Press, 2026	+36% when AI explains reasoning
Trust erosion (no echo)	Francis Press, 2026	−14% when AI skips confirmation

What this demonstrates as a designer: I can translate behavioral evidence into concrete product mechanics, design AI-assisted flows that balance guidance with user agency, and argue for counterintuitive decisions when the research supports them.

Learnings

The hardest design work on this project was subtraction. Fewer questions per turn. Fewer filters at entry. Less information on the card surface. Every cut was intentional and evidence-backed.

What worked: Leading with behavioral evidence made every stakeholder conversation faster. When a design decision has a citation, it stops being a matter of opinion.

What was hardest: Getting the conditional CTA timing right. Knowing when to surface the result count badge required multiple prototype testing rounds before echo-confirmation landed as the right trigger.

What I’d do differently: I would instrument the chat screen from day one — specifically, drop-off data by question position. The serial questioning rule is supported by research, but I’d want production signal before committing to it at scale.

What I’d explore next: I want to understand how people’s confidence changes when an AI brings genuine domain knowledge into the conversation — not just structured questions, but real insight. Markerr has deep rent and market data. I’d want to test whether a renter who gets neighborhood-level context mid-conversation (“rents in this area have dropped 8% this quarter”) makes more confident decisions — and whether that kind of signal reduces the re-search loops we designed around.

The best AI UX is invisible friction removal. The interface should feel like talking to someone who already knows you — not like interacting with a model that’s waiting for the right prompt.