Methodology
Dealers know what buyers actually pay. For a long time, buyers didn't. Here's exactly how CarVeil builds that picture — and where it honestly can't yet.
Data sources
CarVeil pulls live vehicle listings from MarketCheck, with Auto.dev as a fallback source. These aren't static snapshots — benchmark data is refreshed regularly so the prices you see reflect what's actually on the market right now, not what was listed six months ago.
A benchmark is built from comparable active listings: same make, model, year, and condition. When enough listings exist, we compute a range — what the typical buyer encounters at the low end, the median, and the high end of the market.
Asking prices are not the same as what buyers pay. As closed-deal data accumulates, CarVeil will surface what buyers actually signed for — a meaningfully stronger signal. That engine is under active development.
Comparables
A comparable listing — a "comp" — matches your vehicle on four dimensions: make, model, year, and condition (new / CPO / used). Trim is used when enough trim-level data exists; otherwise CarVeil falls back to model-level benchmarks and says so.
Geography matters less than people expect for most vehicles — national supply sets the floor on what dealers can charge, even locally. We use a national dataset by default and flag when local supply patterns differ meaningfully.
Mileage is not a comp filter — it's factored into the deal score directly. A high-mileage car against a low-mileage benchmark is a weaker deal, and the score reflects that.
Grade suppression
When fewer than 5 comparable listings exist, CarVeil does not produce a grade. The threshold isn't arbitrary — a score built on 2 or 3 listings could be meaningfully wrong, and a confident wrong answer is worse than no answer.
Instead, we surface whatever reference data we do have — the model-level median if trim data is thin, or the national range if local supply is sparse — and explain exactly what's missing. You can set a Watch alert and we'll notify you when enough listings accumulate to grade the deal.
Long-tail and low-volume vehicles (rare trims, less common makes) will hit this threshold more often. That's honest. The alternative — fabricating a grade from insufficient data — is not.
How scores are calculated
The deal score is deterministic: the same inputs always produce the same output. It is computed from a fixed formula in code — comparing the asking price to the market median, adjusting for mileage relative to vehicle age, and accounting for comp count confidence.
Claude (Anthropic's AI) produces the plain-language narrative you read — the "watch outs," the negotiation context, the explanation of what the grade means for your specific vehicle. Claude never computes the score. The two are intentionally separate: AI for explanations, arithmetic for decisions.
This means two things: CarVeil can't be talked into a better grade, and you're not trusting a black box with your negotiation. The score is auditable.
Grading scale
Grades map directly to the score range. There's no curve, no adjustment for vehicle type.
| Grade | Score range | Signal |
|---|---|---|
| A+ | 95–100 | Excellent |
| A | 88–94 | Excellent |
| A− | 80–87 | Good |
| B+ | 72–79 | Good |
| B | 65–71 | Good |
| B− | 58–64 | Fair |
| C | 45–57 | Fair |
| D | 30–44 | Weak |
| F | 0–29 | Poor |
Grade suppressed (shown as —) when fewer than 5 comparable listings are available.
Closed-deal and walk-away data
Every benchmark CarVeil shows today is built from asking prices — what the dealer listed the car for. That's useful context, but it's not what buyers actually paid.
When you record what you paid — or what price you walked away from — that data feeds a separate, stronger benchmark. Closed-deal prices and walk-away prices reveal the gap between listed and final, and that gap is where negotiation actually happens.
This benchmark engine is in active development. As CarVeil users record outcomes, the platform will surface what buyers in your position — same vehicle, similar market — actually negotiated. Dealers have always had this data internally. Now buyers will too.
The honesty principle
CarVeil will not produce a grade when the underlying data doesn't support one. This is a product decision, not a technical limitation — it would be easy to generate a number from insufficient inputs. We choose not to.
The same principle applies to AI-generated content. Claude explains what it sees in your deal; it does not fabricate market context that doesn't exist. If negotiation intel is thin because the vehicle is rare, the tool says so.
Trust requires accuracy. Accuracy sometimes requires silence.