AI / LLM

AskZ: designing Zillow's first LLM product

A conversational beta for the highest-stakes consumer decision there is, and the foundation for Zillow's AI Search of today.

RoleDesign lead (and only designer)
Team3 PMs · 3 eng teams · researcher · content strategist
OutcomeGreenlit Zillow's AI copilot; foundation for AI Search
Three AskZ beta screens: the welcome state setting expectations, an answer with inline numbered citations and a sources dropdown, and the report flow behind a thumbs-down.
The shipped beta, in three states. Left: the welcome screen sets honest expectations, including that every question helps the model improve. Center: answers carry inline numbered citations with a Sources dropdown, plus per-response rating. Right: the reporting flow behind a thumbs-down, with "Offensive/discriminatory" deliberately first in the list.

Context

As a senior designer on the Dream and Search team, I was asked in 2024 to lead design for AskZ, a finance-focused chatbot to help people with financial questions on the way to a home, and the company's first LLM product. I was the design lead, partnering with three PMs, three engineering teams, a researcher, and a content strategist, with AI, Search, Mortgage, and Legal leadership all at the table. It shipped on web and mobile web, with webview integrations for iOS and Android.

Real estate is an unforgiving domain for generative AI. A bad answer could lead to a distorted decision about the biggest investment the user could make, or could result in legal action against the company. The quality bar for the LLM's responses had to be set incredibly high.

How I worked

This was a new technology for the whole organization, so I embedded directly with engineering rather than working downstream from them, following their investigation of building on our own LLM versus a GPT integration. I needed to understand what the model could actually do, because those properties are the design material. I also built empathy the unglamorous way: by becoming a heavy LLM user myself (for work, for personal questions, even recipes) until I had instincts for where these systems were delightful and where they were just plain annoying.

Before any interface work, I ran a collaboration session to align the team on ethics: we combined industry-leading AI values with Zillow's core values and our design principles, landing on four commitments: fairness and integrity, inclusiveness, accountability, and transparency, with fair housing compliance woven through all of them. I also brought my content strategist in from day one, because in a product where the interface is mostly language, content strategy is half the design team.

The four AskZ AI principles — fairness and integrity, inclusiveness, accountability, transparency — each paired with the Zillow core value it grew from.
The principles, as workshopped. Each commitment paired with the Zillow core value it grew from. Fairness and integrity lead, with fair housing compliance woven through all of them. These became acceptance criteria for design decisions.

Two directions, three weeks

Leadership gave us a deadline of just a few weeks for design direction, so I presented two: a traditional chatbot pattern, and an emerging AI-native UI. We chose the latter, stripped to what one Microsoft designer memorably called "deceptively simple": a clean, brand-aligned conversational interface where responses stay clear and concise, cutting cognitive load while keeping warmth. Simple to look at; very hard to keep simple.

Two design directions side by side: a traditional chatbot with inline entry and suggested-question chips, and an AI-native pattern with a floating entry and full-screen welcome.
The two directions, as presented to leadership. Left: the traditional chatbot, with an inline entry point and suggested-question chips. Right: the emerging AI pattern, with a floating entry and full-screen welcome with expectation-setting and legal slots.

The transparency principles showed up as concrete patterns, visible in the beta screens at the top of this page: the product clearly disclosed that you were talking to an AI, acknowledged that answers could be wrong with avenues to verify, named its sources explicitly, and was upfront about data use. In 2024 these weren't yet table stakes; we assembled them from first principles and early industry practice.

Requirements alignment

The team worked together to create user stories and priorities — this is a sample of the spreadsheets we were creating together and referencing.

A user stories spreadsheet: rows of P0, P1 and P2 priorities in an as-a / I-want / so-that-I-can format, with a notes column. Entries cover chat ingress placement, session memory, backing out of the chat, understanding the bot's boundaries, error confirmations for unanswerable questions, an accuracy disclaimer, and a feedback prompt. Most rows are checked off.
User stories, priced and prioritised together. Written as P0 / P1 / P2 with the team rather than handed over: the notes column is where the real negotiation happened — which errors needed distinct states, what "memory" meant per session, and where the accuracy disclaimer had to sit. Several of the P0 rows here become the failure states further down this page.

Fair housing: where AI safety got real

If an LLM answers the wrong question about neighborhoods, it can walk straight into illegal steering: influencing where people look based on characteristics protected under the Fair Housing Act. Zillow's own research found that a majority of Americans report experiencing housing discrimination in their lifetime. The team spent a great deal of the project weaving fair housing guardrails into the experience, and then developed a Fair Housing Classifier to detect noncompliant language, which Zillow later open-sourced.

I was in the room as VPs and directors worked through why these guardrails had to exist, and I could contribute to the conversation meaningfully because I'd been a real estate agent. Fair housing law was a professional obligation I'd personally carried! I was then tasked to help train the model and do some "red teaming" — I knew what questions could be asked that would likely trigger inappropriate responses, because I've been asked those questions in real life. Being able to advise the room and help train the model from that side of the table is the contribution I'm proudest of here.

In the press

Zillow open-sourced the Fair Housing Classifier in 2024 so the wider industry could build on it. The release drew coverage in trade press, and the National Fair Housing Alliance's chief responsible AI officer publicly praised the open-source approach as a precedent for responsible innovation. → Zillow announcement · HousingWire coverage

Designing the failure states

An AI product is defined by its worst answer, not its best, so we treated the unhappy paths as core product surface:

  • Unknown answers. When the model didn't know, it said so, and when the question warranted a human, we linked directly to customer service rather than letting the model improvise.
  • Policy violations. Distinct, carefully written states for questions the product wouldn't answer, designed to preserve dignity and offer a path forward.
  • Time-outs and lost connections. Mundane failures, designed with the same care, because to the user, a hung response and a wrong response erode the same trust budget.
  • Grounding every answer. Responses signaled their sources with links to Zillow's own articles and listings, so people could verify rather than take the model's word for it.
The AskZ user flow diagram, with the unknown-answer branch drawn at full weight: error message, suggested new questions, and a customer service escalation annotated on the node.
The flow, with the failure branch drawn at full weight. The unknown-answer path gets equal structural treatment: a specific error message, suggested new questions, and a customer service escalation annotated directly on the node. Negative feedback routes to recovery, not a dead end.

How we tested: internal first, then the real world!

Our primary launch goal was risk mitigation: get a high-quality experience in front of the broader public, not a flashy one. So we tested in phases. First, an internal MVP: we recruited across Zillow, had colleagues use the test environment, leave feedback on responses, and complete a survey, with our researcher synthesizing results. We learned a lot and shipped quick fixes before any external user saw the product.

Then a structured public beta: 10 days, 75 external users with the beta activated on their real Zillow accounts, measured on overall experience, relevancy, perceived latency, ease of use, value, and likelihood to use, followed by a six-person focus group to go deeper than the numbers.

Overall experience
76% of testers had a positive overall experience, and it was especially helpful for first-time and early-stage home buyers.
Relevancy
83% found AskZ's responses relevant, though some thought they were sometimes inaccurate and too generic.
Perceived latency
70% felt AskZ was fast enough. The general impression: "could have been faster but not extremely slow."
Ease of use
97% found AskZ easy to use — our strongest result of the round.
Value & likelihood to use
59% found AskZ valuable in their home buying journey and 58% said they'd likely use it. Our weakest scores, and improving accuracy would lift both.

The clearest improvement path out of the beta was that responses were too text-heavy and took a lot of effort to read. That pointed toward embedded objects, data visualizations and calculators as the way to make answers genuinely useful. Testers loved what worked: 24/7 availability and fast answers to simple questions. That signal is ultimately what unlocked the investment that followed.

Outcome

AskZ cleared the bar it was built to clear! BETA testing results ultimately led to leadership greenlighting the investment in a full-service Zillow chatbot, now being evolved into a sophisticated AI copilot.

Three screens from the copilot vision: a search feed with rental listings, a voice-input state showing a waveform and 'Listening…', and a conversation where a shopper asks in plain language for homes with open kitchens, a second bedroom for an office and outdoor space, and gets matching results back.
Where it went next: the copilot meets Search. The beta's finance-only scope becomes an assistant that sits inside search itself — asked in plain language, out loud, with results returned as listings rather than paragraphs. (Prototype screens.)

The fair housing guardrail work grew beyond our product entirely. Out of the combined effort, the engineers on my team created the Fair Housing Classifier, which Zillow open-sourced so the rest of the industry could build on it.

The published paper 'Navigating Fair Housing Guardrails in LLMs', filed under AI, Machine Learning & Research, dated January 16 2024. An illustration shows a robot reading a listing on a phone while a shopper holds another, with an LLM cog feeding into a shield and check mark. Authors: Ondrej Linda, Eric Ringger, Gabe Arnson, Anusha Bagalkotkar and Aveek Karmakar.
The paper that came out of it. "Navigating Fair Housing Guardrails in LLMs," published January 2024 — the engineering write-up behind the open-sourced classifier.
What I'd do differently

I committed to a three-week turnaround for design direction and I sacrificed quality for speed — I should have quoted a realistic timeline. It's a mistake I only needed to make once: I now scope AI design work with plenty of padding to understand the underlying technology and contribute to it! More than ever, I felt the importance of a close relationship with engineering and science. This was so complex at the time, and they were a lifesaver through the whole process.