Utah Dealers.ai
Sign in
Methodology

How the 7-Star Trust Score works

Every star is a transparent equation — not an opinion. The score is computed for every licensed Utah dealer from verified public data, by a published, deterministic formula. AI reads the inputs; the math decides the number. It is never for sale.

★ = Tier( Σ wₚ · Percentile( Shrink( WeightedMean(reviewsₚ) ) ) − z·σ/√N )

The four pillars we weigh

Published weight ratio 35·30·20·15. Pillars without real data yet are imputed at a conservative class floor and shown honestly as “Data pending” — they cap the score until real evidence arrives.

Sales Experience
Verified sales reviews (Google today; per-review sources as they land)
35%
Service Experience
Service reviews · quality-of-repair · wait-time & comeback signals
30%
Employee & Culture
Licensed employer-review data · first-party eNPS — data pending, conservative floor applies
20%
Complaint Resolution
State consumer-protection + BBB closure records — data pending, conservative floor applies
15%

The seven tiers

Anchored to the real Utah distribution — so the median dealer is a 4, and a 7 is genuinely rare.

7
ExceptionalGated — rare by design
92–100
6
Excellent
80–91
5
Strong
66–79
4
SolidUtah median lands here
50–65
3
Mixed
35–49
2
Poor
20–34
1
Critical
< 20

The math — and the ethic behind each step

①Weight each review
w = R(t) · V(v) · c · Λ(r)
Why: Recent, verified, real experiences count most — and a harmful one weighs more than a routine good one.
②Weighted average
x̄ = Σ(w·r) / Σw
Why: A dealer's record — not a loud minority.
③Shrink to a fair prior
μ = (x̄·n + 3.8·C) / (n + C)
Why: Nobody is crucified on two reviews.
④Normalize vs a locked Utah cohort
P = percentile-rank(μ, cohortᵣₑբ)
Why: Honest relative truth against a fixed reference — bands can't drift quietly.
⑤Apply the fair pillar mask
N/A → dropped · missing → 25th-pctile class floor
Why: Never punish a dealer for a service they don't offer; hiding data can't inflate a score.
⑥Combine the pillars
Q = Σ(wₚ · Pₚ) / Σ wₚ
Why: Published weights — anyone can reverse-engineer it.
⑦Publish the lower bound
Q_LB = Q − 1.645·σₘₐₓ / √N
Why: A one-sided 95% bound with worst-case variance — a few lucky reviews can't outrank a proven record.
⑧Map to a tier
Q_LB → ★ 1–7
Why: Most are a 4. A 7 is earned — and requires 25+ verified signals with a tight interval.

The firewall

A dealer can pay $499/yr to claim and verify their profile — that confirms their identity and active Utah license. It never changes their Trust Score. Payment never reaches the equation, and every score change is written to a hash-chained ledger that cannot be quietly altered.

Methodology changelog

Every change to the math is versioned, published here, and recorded in the audit ledger before scores recompute.

B-0.2July 2026

Conservative one-sided 95% lower bound (Bhatia–Davis worst-case variance) replaces the fixed-σ interval · missing-pillar floor is now the 25th percentile of the dealer's class · 7★ gate enforced (≥25 verified signals, interval ≤0.5★) · percentile cohort locked to an audited reference snapshot · every score's exact inputs are hashed into the ledger.

B-0.1June 2026

Launch methodology: Google aggregate as the sales+service proxy, Bayesian shrinkage to the Utah prior, percentile normalization, lower-bound publishing, missing-pillar floors.

Methodology version B-0.2. Disagree with a score? Dispute it →