Desktop Reviews & topic visibility
Pipeline topics; sort unchanged; disclosure before interaction. Ops monitor recurring themes without per-review manual tagging.
Reducing review overload through AI-powered apartment insights.
ApartmentRatings property pages held hundreds of reviews. Renters wanted to spot recurring issues fast—but newest-first feeds buried patterns and keyword search failed when wording varied. I designed semantic topic chips inside the existing Reviews tab so renters could navigate by pattern, not search for it.
A product design case study about using AI for navigation, not narration—designing for renter trust, scaled content, and cross-functional constraints.
ResultThe feature shipped as AI Trending Topics and introduced a scalable framework for surfacing AI-generated review insights while preserving user trust and existing review workflows.
Product Design · AI-powered discovery · Marketplace
Overview
On high-volume property pages, renters struggled to spot recurring issues across hundreds of reviews. I introduced semantic topic labels inside the existing Reviews tab so renters could navigate by pattern—not search for it. Resident feedback stayed the source of truth; AI handled discovery, not interpretation.
Business Problem
ApartmentRatings owned one of the largest renter-review corpora in the marketplace—but volume was working against the product. More reviews meant slower decisions and higher bounce on the pages with the most ad inventory.
Why this mattered
- High-review properties had the most inventory and weakest discovery.
- Renters bounced before reaching the contact CTA on dense pages.
- Manual theme tagging by ops didn’t scale.
- AI was a strategic bet—but couldn’t weaken Verified Reviews.
The opportunity
If renters could see which themes recurred at a glance, they could judge dealbreakers in seconds—not twenty minutes of scrolling.
Solvable with classification, not generative copy.
User Problem
Renters weren’t looking for more information. They were looking for a way through the information they already had.
What renters experienced
- Information overload on properties with 50+ reviews.
- Search failed when wording varied (“thin walls” vs. “noise”).
- Newest-first feeds buried the patterns that mattered.
- They didn’t know what to search for.
“I don't have time to read 100 reviews—I need to know if noise or maintenance keeps coming up.”Renter feedback, usability synthesis
Constraints
The shape of the solution was set as much by what we couldn’t do as by what we wanted to ship.
Product & legal
- Verified Reviews placement and hierarchy could not change.
- No AI paraphrase or summary that renters couldn’t audit.
- AI disclosure visible on first paint, before any interaction.
Engineering & system
- Topic labels from a classification pipeline—not real-time inference.
- Low-volume properties needed graceful empty states.
- Must integrate with the existing search and sort architecture.
- Mobile parity required—no forked mobile feature.
- Ship in one quarter alongside Verified Reviews maintenance.
What We Learned
Research paired Reviews-tab analytics with moderated sessions on dense pages. Each finding had to point to a concrete design implication—otherwise it didn’t earn a slot in scope.
Research insight
Renters scanned for patterns before reading any individual review. In moderated tasks on dense pages, they paused on the label row and used counts to decide what to open.
Design response
Topics had to live above the feed as the primary hierarchy, with counts as a first-class signal—not metadata.
Product outcomeA fixed topic row earned the visual weight that the feed used to consume.
Research insight
Renters distrusted AI when it summarized other residents. In pattern audits, they skipped summary blocks and went straight to original cards.
Design response
AI could organize content—not speak for residents. No summary modules; disclosure on first paint.
Product outcomeVerified Reviews stayed the source of truth. We shipped navigation, not narration.
Research insight
Keyword search failed when wording varied. “Thin walls” and “noise” returned different result sets; renters bounced after the first miss.
Design response
Discovery had to be topic-led. Search refines within a topic; it doesn’t start exploration.
Product outcomeSemantic labels group varied phrasing under one chip. Search dead-ends dropped in usability runs.
Research insight
Mobile was where the decision happened. Most high-intent traffic came from mobile, and mobile users scrolled less and exited sooner on dense pages.
Design response
Mobile parity wasn’t polish—it was the launch surface. Mobile led every design round.
Product outcomeOne filter model across breakpoints. No forked mobile feature.
Journey Map
Before exploring solutions, I mapped the renter journey to identify where cognitive load was highest and where AI could provide meaningful assistance without taking over the renter’s judgment. The map made the opportunity precise: Discover and Evaluate were where renters disengaged—not Decide.
- 01
Discover
FeelingCurious · distracted
GoalIs this topic worth attention?
ActionsScan trending lists and labels.
FrictionFlat hierarchy. Every topic reads equally important.
Design opportunityClear ranking. Stronger labels and visual priority.
- 02
Evaluate
FeelingOverwhelmed · skimming
GoalGrasp the conversation fast.
ActionsOpen thread; skim top responses.
FrictionWall of text. Repeated takes spike cognitive load.
Design opportunityScannable grouping. Tighter spacing and content structure.
- 03
Compare
FeelingSkeptical · low trust
GoalSeparate signal from noise.
ActionsWeigh engagement and contributor quality.
FrictionLow-trust mix. Hard to spot credible insights.
Design opportunityCredibility cues. Surface engagement and relevance.
- 04
Decide
FeelingFatigued · indecisive
GoalForm a view without exhaustion.
ActionsScroll for diverse perspectives.
FrictionBuried viewpoints. High-signal takes lost in repetition.
Design opportunityPrioritize diversity. Elevate distinct, high-signal replies.
- 05
Contact / Convert
FeelingDrained · likely to bounce
GoalEngage—or move on with clarity.
ActionsContribute, bookmark, or leave.
FrictionValue lag. Effort exceeds payoff before meaningful insight.
Design opportunityProgressive disclosure. Less load; faster path to value.
Key opportunity
Renters had the information they needed—competing signals made relevance hard to recognize at a glance. AI could navigate, not narrate.
Design Principles
Five principles framed the decisions. They were the contract between design, product, and engineering—what we’d ship, what we’d refuse, and how we’d evaluate a proposed change later.
Reduce cognitive load
Surface the few themes that recur most. Let counts do the prioritization work.
Support exploration, not just search
Renters didn’t know what to look for. Topics had to be browseable entry points.
Build trust through hierarchy
AI organizes; residents speak. Original reviews stayed the proof.
Preserve user control
One filter at a time. A visible reset gets back to the full feed in one tap.
Scale for future AI features
Chips, disclosure, and filter state designed as a reusable pattern—not a one-off.
Design Decisions
The harder calls weren’t about visual design—they were tradeoffs between scope, trust, and renter behavior. Each decision was made with Product and Engineering at the table.
One topic filter at a time
Why. Combining topics broke the mental model and parity with site search.
Tradeoff. Power users lost topic intersection.
Outcome. Predictable filter behavior that shipped on time.
No AI summary module
Why. Renters distrusted paraphrase; Legal required Verified Reviews stay the auditable source.
Tradeoff. Less visible “AI surface area” at launch.
Outcome. A defensible launch and a trust model the team can extend later.
Topics inside the existing Reviews tab
Why. A new tab would have signaled “AI experiment” and split ops, analytics, and content models.
Tradeoff. Less marketable; competed for vertical space in a busy tab.
Outcome. One tab, one model, mobile/desktop parity.
Search refines a topic, not the inverse
Why. Search-then-filter produced empty result sets and high abandonment.
Tradeoff. Renters who knew the exact phrase had to pick a topic first.
Outcome. Fewer dead ends; clearer recovery when filters over-narrowed.
Solution
V1 shipped as filter states inside the Reviews tab: labeled topics, active chips with counts, in-topic search, and a reset—plus desktop visibility for pipeline-driven topics. Hierarchy and disclosure were identical across breakpoints.
Alternatives considered
- AI summary card above the feed — rejected for trust.
- Dedicated “Insights” tab — rejected to avoid forking ops and analytics.
- Multi-topic combinations — deferred; tested poorly.
- Free-text “Ask this property” — out of scope for V1.
Why we shipped this
- Most decisive lift on the Discover and Evaluate bottlenecks.
- No compromise to Verified Reviews hierarchy.
- Reusable pattern for future AI features.
- Achievable in the engineering window without a new AI surface.
Topic labels as entry points
Active chip and count state make the current filter obvious. Renters validate labels by reading the cards—not by trusting the label text alone.
In-topic search & recovery
Search applies only after a topic is selected—matching how renters refine, not start, discovery.
Mobile parity
Same hierarchy on mobile: topics → filter → cards. One filter state, one disclosure pattern, one hierarchy. No forked mobile feature.
Impact
We tracked Reviews-tab behavior and moderated follow-ups—not generic “AI engagement.” Signals below are what the team observed; anything we couldn’t measure is framed as an expected product outcome, not a claim.
Faster pattern recognition
In follow-up usability sessions, renters reached a themed review set in fewer interactions than unfiltered browse.
Trust signals intact
Verified Reviews placement and sort unchanged through launch QA. Disclosure visible on first paint, all breakpoints.
Ops-ready labeling
Pipeline-driven topics replaced manual theme tracking in partner demos and support walkthroughs.
Scalable AI pattern
Chips, disclosure, and topic-first filtering became a reusable framework future AI features extend.
Mobile/desktop consistency
One filter model across breakpoints. Removed a forking risk between mobile and desktop work.
Topic module engagement
Trending topic chips became a measurable interaction surface on high-volume property pages.
Reflection
The most strategic AI decision wasn’t a model choice—it was the choice not to summarize. Saying no to a summary block early protected the trust the rest of the product depends on.
Working cross-functionally surfaced how much of “AI design” is actually system design. The interesting tradeoffs lived between Product, Engineering, and Legal. Getting those constraints aligned was the real interaction-design work.
I’d push earlier on mobile next time. Mobile parity was a goal from day one, but the strongest scan signals came from mobile tests—I’d open every round there and let desktop catch up.
The framing that changed everything: ship navigation, not narration. It defined what we built, what we deferred, and how the team talks about AI on the product today.