Buying blind: $20k decisions, no test drive
ACV Auctions is a digital wholesale car marketplace. Dealers buy vehicles they've never touched, from sellers they've never met, on a time limit. The average vehicle sells for $18,000–$24,000. The goal isn't to make decisions faster, it's to make them more confident — for both sides. Misrepresent a vehicle and buyers dispute it. Underrepresent one and sellers take their inventory to Manheim.
The VDP is where that decision happens. Everything a dealer needs to bid — condition, photos, inspection data, pricing — lives in one place. Get it wrong and they don't bid. Get it right and they come back.
The VDP was making good vehicles look bad.
The VDP was a data dump, not a narrative on whether to buy or walk away. Every finding at the same visual weight, no positive framing, no baseline for what was normal. The design had evolved reactively: disclose everything, protect against arbitration. Over time, features were added, not designed. It looked thorough. It eroded buyer confidence.
Data without a story is just noise.
A clean, sellable unit presented identically to one with structural damage. Buyers had no way to distinguish "this is normal" from "this is a problem." The interface offered data. It offered no judgment.
Bids dropped, abandonment increased, sellers moved to Manheim and OpenLane. Good inventory was systematically undervalued because every vehicle was presented like a risk.
This wasn't just a UX issue. It was a marketplace problem.
How should condition be structured?
Why the tiers exist
The same disclosure list reads completely differently depending on the vehicle. On a 120k-mile truck, 20 findings mostly mean Expected Wear — excluded from arbitration by policy. On a 3-year-old SUV, five findings hit differently.
Highlights
Verified positives — not just the absence of problems. Single owner, below-market mileage, clean engine audio, high-demand spec. If a vehicle has something going for it, buyers should know before they bid.
Summit trim, 4WD, Panoramic Sunroof, UConnect 5 Nav — matches current buyer demand for mid-size SUVs
No abnormal engine noise detected across all RPM ranges
20,150 mi — 18% below average for a 2022 Grand Cherokee Summit
One owner since new — confirmed via CARFAX
Known Recon
Confirmed issues that will cost money. Visible damage, mechanical failures, things an inspector could see and price. Because the cost is known, our AI surfaces repair estimates directly in the listing. No math required.
Cracked bumper, grille, trim + loose mirror casing.
Cracked Head Lamp
Requires Diagnosis
Something is flagged but the cause isn't confirmed yet — a warning light, an inconsistent reading, something that needs a mechanic to scope. No cost estimate until it's confirmed.
System fault detected.
Monitors incomplete (Fuel, Catalyst, EVAP, O2, EGR/VVT).
Malfunctioning Display Screen
Expected Wear
Normal findings for the age and mileage. Light cosmetic wear, typical interior use. Shown so buyers know what's normal — not as a problem to factor in.
Minor cosmetic imperfections such as light scratches, chips, and small dents are present.
Light seat wear and surface use consistent with normal driving.
Normal wear & tear is excluded from ACV's arbitration policy and cannot be used as grounds for a post-sale claim.
View arb policyThe hardest part: legal alignment
What helps a dealer act — a number, or a sentence?
Why scores failed
Condition confidence scores are the industry default. But in dealer research, buyers consistently couldn't explain the difference between a vehicle with a condition rating of 4.1 and one rated 4.3. The opportunity wasn't to produce a better score. It was to replace the score entirely with something a dealer could act on in seconds.
Scores compress everything a dealer needs to know into a number that strips the context required to act on it. The narrative restores that context — and does five things a score never can:
Single owner, 20,150 mi — low for a 2022. Known recon: front-end damage and cracked head lamp — estimated $1,650. Three items require diagnosis: ABS fault, emissions readiness failures, and malfunctioning display screen — keep these in mind when you bid. Wear consistent with normal use.
The trust risk
The model only works with what's been disclosed — it never fills gaps with inference. What it can't confirm, it doesn't say. The dealer still decides.
"What can we confidently say?" was a shared question between design and engineering — set before the model was built, not negotiated after the fact.
The old VDP was a data dump — features bolted on over time, never overhauled. How do you turn that into something a dealer can act on in seconds?
Five questions, zero answers
The old VDP wasn't built to answer the questions dealers actually ask. Information was scattered across the page with no hierarchy or intent. Announcements read like a flat list of issues — no structure, no signal, no way to quickly understand what mattered.
No market context meant dealers were guessing whether a vehicle would move in their lot. No seller data meant every transaction started from zero — no track record, no reason to come back.
And when it came to cost and profit, the workflow broke entirely — dealers had to leave the platform just to run the numbers.
1FAFP31N57FML5AGU
System fault detected.
Malfunctioning display screen
Cracked bumper, grille, trim + loose mirror casing.
Cracked head lamp
Paint work on bumper, fender, hood, liftgate
Irregular wear front-left and rear-right
What we gained, gave up, and are still calibrating
A page that thinks like a dealer
Restructuring the IA around five decision questions meant every section had to earn its position. The page asks the same questions a dealer does, in the same order. They don't have to reverse-engineer what matters — it's already organized that way.
The simplicity of showing less
Adding seller transparency, market context, and a profit calculator created more surface area for data freshness issues, edge cases, and support burden. The old page was easier to maintain precisely because it surfaced so little.
A trust loop between buyers and sellers
Surfacing seller history, arbitration rate, and purchase history creates a feedback loop that rewards good sellers and gives buyers a reason to return. The data already existed — we just finally surfaced it where the decision was being made.
Speed to ship
Each new component had a dependency — a data pipeline that didn't exist, a legal review, an engineering estimate that kept growing. A cosmetic refresh would have shipped in weeks. This took months and created real tension while an MVP was already in progress.
Pre-filled estimates vs. false precision
The profit calculator defaults to pre-filled numbers — inspection recon, transport, market retail. Convenient, but dealers may anchor on those figures even when they're imprecise for their market. The tension between useful defaults and misleading specificity is ongoing.
Seller transparency vs. seller sensitivity
Showing arbitration rates and trust signals is valuable to buyers but uncomfortable for sellers. The line between transparency that builds marketplace trust and data that feels like public judgment is not a line you draw once — it requires ongoing calibration.
Every new data surface required a live, reliable data source. Market fit needs real demand signals. Seller profile needs arbitration history. The profit calculator needs live pricing comps. The design was only as good as the underlying data infrastructure — which meant some features shipped with caveats, and the full vision required backend work running in parallel with design.
What changed
Initial 90-day data post-launch, measured against the prior year equivalent period. Some metrics are directional — the system is still rolling out across surfaces — but the signals are consistent enough to draw conclusions.
The arbitration reduction is the metric I weight most — it's the one most directly tied to the condition model redesign, and the one with the clearest business cost behind it. Bid conversion and confidence scores improved alongside it, but I'd want another 90-day cycle before claiming full causality on those numbers.
The metrics are one part of it. The bigger thing: the condition model, the trust framework, and the modular system built here are what AI features on the platform now plug into. We weren't just redesigning a page — we were building the foundation.
What this project taught me
Trust is not about more data — it's about structure. The buyers who felt most uncertain weren't the ones with the least information — they were the ones who couldn't make sense of what they had. How you organize data matters more than how much of it you show.
Legal alignment is design work. The condition language guardrail — factual and verifiable, not interpretive — only exists because I treated the legal conversation as a design problem. Framing, evidence, iteration. The specific outcome: a platform-wide standard that distinguishes what ACV can assert from what it can only imply. If I'd treated legal as a blocker instead of a collaborator, we'd have shipped a worse product faster.
AI has a ceiling — and design has to set it. The Smart Summary works when it synthesizes; it fails when it asserts. The guardrails aren't UI copy — they're model-level constraints that required design and engineering to define together. "What can we confidently say?" is a joint question.
Systems take longer and pay back more. The single-surface fix would have shipped two months earlier. The universal system shipped two months later and eliminated 10+ versions of compounding design and engineering debt. That return wasn't obvious upfront — it had to be argued for.
Building while defining creates rework. The MVP was in progress while the system vision was still being established. Some engineering work had to be redone. I'd push to freeze MVP scope earlier and invest more in system definition before build begins — even when that feels slower in the moment.
Design artifacts
These artifacts represent the thinking behind the product, not just the output. Each one was a working tool during the project — used to drive alignment, resolve debates, and establish shared language across design, legal, and engineering. Visuals pending permission from ACV.