Methodology

Handling missing data

This is the part that matters. A component whose inputs are unavailable is removed from the calculation, not scored zero. The final score is the weighted mean of the components that could actually be computed.

Scoring a missing input as zero would be the easy implementation and a badly misleading one — a car with no asking price entered would look like a terrible deal rather than an unassessed one.

Because the denominator shrinks with missing data, the score is always paired with a confidence label derived from the share of total weight that was available: 80% or more is Good, 50% to 80% is Partial, below 50% is Limited.

Score bands

• 85–100 — Excellent Deal
• 70–84 — Good Deal
• 55–69 — Fair
• 40–54 — Questionable
• 0–39 — Poor Deal / High Risk

What the Deal Score is not

It is a presentation heuristic for comparing listings against one another. It has not been validated against transaction outcomes, and VINLytix makes no claim that it is statistically predictive. The bands above are user-experience thresholds chosen for legibility, not empirically derived cut-offs. Every report shows the full component breakdown so you can judge the reasoning rather than trusting the number.

Data Coverage

A separate figure from 0 to 100, answering a different question: how much of the relevant vehicle record VINLytix was able to verify. The domain list and its weights are fixed, so coverage cannot be inflated by shortening the checklist.

Two domains — recall repair completion and full history — are currently unreachable for every report, which places a hard ceiling of 75% on Data Coverage. That ceiling is deliberate and is not hidden. Without a valuation provider connected, the practical ceiling is 55%.

Deal Score and Data Coverage are independent. A vehicle can score 88 on 45% coverage. That combination is not a contradiction, it is the honest answer: what we could see looks good, and we could not see much.

Mileage analysis

Expected mileage accumulates on a tapered schedule, because annual mileage genuinely falls as a vehicle ages:

• Years 1–5 — 13,500 miles per year, close to the average annual mileage per US light-duty vehicle reported by the Federal Highway Administration
• Years 6–10 — 11,000 miles per year
• Years 11 and beyond — 8,000 miles per year

A flat rate would compound its error until it became absurd: 13,500 a year across 23 years predicts over 310,000 miles, so a 168,000-mile 2003 sedan would be scored “below average” against a figure almost no car of that age reaches. The taper keeps the comparison meaningful across the whole age range.

Age is the current calendar year minus the model year, with a floor of half a year so a current-model-year vehicle does not produce a division by zero.

Verdicts are set on the ratio of actual to expected:

• Below Average — under 0.75x
• Average — 0.75x to 1.15x
• Above Average — 1.15x to 1.5x
• Very High — over 1.5x

This compares one odometer reading against a population average. It is not odometer fraud detection, which requires a history of recorded readings from a licensed provider. VINLytix will never suggest a rollback on the basis of this calculation.

Valuation Confidence

A third measure, separate from both of the above, and the three answer different questions:

• Deal Score — how attractive the deal looks
• Data Coverage — how much of the vehicle record was verifiable
• Valuation Confidence — how solid the price estimate is

A car can score 85, on 70% coverage, from a valuation we have Low confidence in. Those are three separate facts and the report shows all three.

Confidence is an additive score over five signals:

• Comparable listings (30) — 200+ scores full; 50–199 scores 22; 10–49 scores 12; fewer scores 4. Zero if the provider does not report a count.
• Trim match (20) — full when the provider matched the exact trim, 6 when it explicitly could not, 10 when it does not say.
• Mileage input (20) — full when an odometer reading was supplied, zero otherwise. Mileage is one of the largest price factors.
• Local market (15) — full when the value is anchored to a local market, 8 when a ZIP was given but the figure is national, zero when no ZIP was supplied.
• Data recency (15) — full at 7 days or fresher, 10 up to 30 days, 3 beyond that, 5 when the provider does not publish a dataset date.

Bands: High 75+, Medium 50–74, Low 25–49, Insufficient Data below 25. A provider that reports its own low confidence overrides an otherwise optimistic total — it knows things about its model that we cannot see.

When confidence is Insufficient Data, the report still classifies the asking price but withholds the negotiation range. A dollar figure presented as advice implies a valuation solid enough to negotiate against, and lending that precision to a number we have already said we do not trust would be the wrong kind of confident.

The VINLytix AI Market Estimate

VINLytix does not have a secret dealer price book, and it does not ask a language model what a car is worth. It estimates market pricing the way a careful buyer would if they had all afternoon: by finding vehicles like yours that are actually for sale right now, discarding the ones that are not really comparable, and taking the middle of what is left.

The difference between that and a guess is evidence. Every estimate on a report is accompanied by the listings it was computed from, with links. If the number looks wrong to you, you can go and check.

How comparable vehicles are found

The search starts as narrow as your VIN allows and widens only when the narrow search did not find enough:

• Level 1 — exact year, make, model and trim, within a local radius.
• Level 2 — exact year, make and model. Trim relaxed.
• Level 3 — within one model year, wider area.

Make and model are never relaxed at any level. A thin search for an Accord does not quietly start pricing Camrys. The radius expands through 25, 50, 100 and 250 miles before falling back to a national search, and the report always states which level and which radius produced the figure — twelve listings at Level 1 supports a very different claim from twelve at Level 3.

If no ZIP code was supplied there is no local market to search, so the result is labelled a National market estimate and carries lower confidence.

What gets excluded

Most listings a search returns are not usable comparables. These are removed before any arithmetic, and the report says how many went and why:

• Monthly payments and lease rates advertised where a sale price should be
• Placeholder prices — a published figure of a dollar or two meaning “call for price”
• Salvage, flood, rebuilt and otherwise branded vehicles
• Parts-only and non-running cars
• Auction bids, which are not asking prices
• New vehicles
• A different make, model, or materially different trim
• Model years outside the range the search level permits

The lease filter matters more than it sounds. A monthly payment accepted as a car price does not look like an error in the output — it looks like a bargain.

Duplicates

The same physical car is routinely advertised on the dealer’s own site and on two marketplaces. Counting it three times would both inflate the comparable count that drives confidence and let one unrepresentative vehicle move the estimate three times as far as it should. Duplicates are detected by published VIN, by canonical URL, by shared photograph, and finally by matching vehicle, mileage and price at the same seller.

How the price is calculated

The headline figure is a weighted median of the accepted asking prices, after removing statistical outliers.

• Median, not average. Three optimistic dealership listings can drag an average anywhere. They can only move a median to the next car along.
• Weighted. A near-identical car three miles away is better evidence than a different trim two states over, and the arithmetic should say so. Weights come from a similarity score over model year, mileage, trim, distance, listing recency and how complete the listing data was.
• Outliers removed. Prices beyond 1.5 times the interquartile range are excluded — the extra digit a dealer typed by mistake. Below six listings this is skipped, because with four prices one genuinely expensive car looks like an outlier and removing it is just discarding evidence.
• Typical asking range is the interquartile band: the middle half of what these cars are advertised at.

All of that arithmetic is ordinary deterministic code. No language model computes any figure on a VINLytix report. An AI may help read a messy listing title or normalise a creatively-written trim; it never produces, adjusts or checks a price.

Mileage and location

There is no universal cents-per-mile constant here, because there is no universal cents-per-mile constant in the world — it differs by segment, age and model. If the retrieved listings contain enough of a mileage spread to measure a price-per-mile relationship, that measured rate is used and the report says so. If they do not, no mileage adjustment is applied and confidence drops instead.

The same principle governs geography. Rather than applying invented regional multipliers, the engine prefers local listings and lets the data show what the local market actually is.

When there is no estimate

Three comparable listings is the floor. Below that, the report says “Not enough current market data to produce a reliable estimate” and shows no figure at all. Two listings can be thousands of dollars apart with no way to tell which is typical, and publishing their midpoint would be a fabrication dressed as an analysis.

Confidence above that floor considers the number of comparables, how similar they are, whether trims were stated, whether the search stayed local, how far the search had to widen, how tightly the listings agree on price, whether more than one source corroborates them, and how old they are. Roughly: 11 or more close local listings can reach High; 6 to 10 sits around Medium; 3 to 5 is Low.

The result is an estimate derived from current asking prices. It is not an official valuation, not a certified value, and not a guaranteed sale or purchase price. Asking prices are what sellers want, which is not always what cars sell for.

Where each money figure comes from

Every valuation figure on a report is tagged with its provenance, and the two tags mean genuinely different things:

• Provider data — the pricing provider supplied this number.
• VINLytix estimate — the provider did not supply it, so VINLytix derived it. The calculation is printed underneath the figure, every time.

Where a provider returns only a retail estimate, the derived figures use these conventional rules of thumb for the US used-car market. They are published here precisely so a reader can discount them appropriately — they are not measurements:

• Personal Buyer Target — the retail estimate. A private buyer keeping the car pays about market value.
• Dealer Buy Target — retail less 15%, a conventional allowance for reconditioning, overheads, risk and resale margin. This is not a wholesale quotation. A real trade offer depends on the car’s condition and the buyer’s stock needs.
• Quick Sale Estimate — retail less 8%, the discount typically needed to sell privately in days rather than weeks.
• Suggested Listing Range — retail plus 4–9%, leaving room to negotiate without being filtered out of buyers’ price searches.

A figure VINLytix cannot derive defensibly is shown as unavailable rather than guessed, and a modelled figure is never rendered without its basis.

Why this value?

Where a provider breaks its estimate into factors — mileage, trim, local demand — the report shows them. Where it does not, the section says so. VINLytix does not generate statements like “trim adds $900”: that claim needs data behind it, and inventing a plausible attribution is exactly the kind of confident-sounding fiction this product exists to avoid.

Data freshness

Two dates, kept deliberately apart on every source label:

• Checked — when VINLytix queried the source. Always known.
• Data updated — when the source’s own dataset was last computed. Shown only where the provider publishes it, and omitted rather than guessed otherwise.

Querying a source today does not mean its data was updated today, and the report never implies that it does.

Price classification

Where both an asking price and a market valuation exist, the signed percentage difference against the estimate is classified as:

• Great Deal — 10% or more below the estimate
• Good Deal — 3% to 10% below
• Fair Price — within 3% either way
• Above Market — 3% to 10% above
• Overpriced — more than 10% above

The recommended negotiation range is anchored just below the point estimate, and is never allowed to fall below the pricing provider’s own lower bound — suggesting an offer outside the provider’s stated range would be advice we cannot support.

VIN validation

Every VIN is checked for length, for the letters I, O and Q which never appear in a VIN, and against the ISO 3779 position-9 check digit. The checksum transliterates each character, applies positional weights, and takes the sum modulo 11.

A failed check digit is reported as a warning, not a rejection. The check digit is mandatory for vehicles built to North American specification and optional elsewhere, so legitimately imported vehicles routinely fail it.

Recall campaigns are never a “major” red flag

Campaigns are matched by year, make and model, so the count reflects the vehicle line’s history rather than the condition of the car in front of you. A 2003 sedan can legitimately return two dozen campaigns, nearly all of them repaired years ago under the free-repair obligation.

Treating a high count as a major issue would contradict the caveat printed beside it and would condemn essentially every older vehicle on the strength of data we have already said cannot support that conclusion. Recall findings are therefore capped at needs attention, with the report directing you to the only source that can actually answer the question — a franchised dealer checking the individual VIN.

The verdict

The final verdict combines the Deal Score, the presence of major red flags, and coverage. Any major red flag, or a score below 45, produces High Risk / Poor Value. A score of 70 or above with nothing needing attention produces Worth Considering — but only if Data Coverage reached 50%. Below that, the verdict is capped at Proceed With Caution, because a confident endorsement built on thin information is not a confident endorsement.

The “watch” list is never empty.

Changes to this methodology

These calculations will change as providers are connected and as the scoring is refined. Material changes will be reflected on this page. See which sources are currently active.