Most dealerships do not have a pricing problem or a stocking problem first. They have a vehicle data quality problem that quietly distorts everything downstream. When a VIN is mistyped, a mileage reading is entered in the wrong unit, or the same car exists twice under two stock numbers, every report, valuation and stock decision built on those records inherits the error. The car looks fine on the forecourt. The data behind it does not.
This article treats vehicle records the way a data team treats any critical dataset: as something to be measured, monitored and remediated against explicit rules. It is deliberately distinct from two neighbouring topics. It is not about where a value came from (that is lineage), and it is not about which system is the master copy (that is architecture). It is about the state of the data itself: how complete it is, how accurate it is, how many duplicates hide in it, and what validation should have caught before it ever reached a pricing model or a stock report.
What dirty vehicle data actually looks like
Dirty data is an abstract phrase until you see it in a real stock list. In vehicle records it tends to fall into a small number of recurring patterns, and recognising them by name is the first step to measuring them.
- Duplicates. The same physical car appears more than once, often because it was re-entered after a failed import, or because a trade-in and a stock record were never merged. Duplicates inflate stock counts and double-count value.
- Gaps (incompleteness). Missing fuel type, missing service history flag, missing first-registration date. A model that needs these fields will either skip the car or substitute a default, both of which corrupt the output.
- Bad VINs. A 17-character VIN that is the wrong length, contains the forbidden letters I, O or Q, or fails its check digit. A bad VIN breaks decoding and links the record to the wrong specification.
- Unit and format errors. Mileage in the wrong unit, dates in mixed formats, prices with or without VAT recorded inconsistently. These are the hardest to spot because the value looks reasonable.
- Stale records. A car marked available that sold three weeks ago, or a price that was never updated after a repricing decision.
The six dimensions of vehicle data quality
To manage data quality you have to make it concrete. The standard data-quality dimensions map cleanly onto vehicle records, and each one suggests a specific test you can run.
| Dimension | Question it answers | Vehicle example | A test you can run |
|---|---|---|---|
| Completeness | Are required fields present? | Is fuel type populated? | Share of records with all mandatory fields filled |
| Accuracy | Does the value match reality? | Does mileage match the service record? | Sample against a trusted source |
| Consistency | Does it agree across systems? | Same price in DMS and on the website? | Cross-system field comparison |
| Validity | Does it obey the rules? | Is the VIN 17 valid characters? | Format and check-digit validation |
| Uniqueness | Is each car recorded once? | One record per VIN? | Count of duplicate keys |
| Timeliness | Is it current? | Is the status up to date? | Age since last update against a threshold |
You do not need all six perfect on day one. Pick the two or three that hurt most - for many used-car operations that is uniqueness, validity and completeness - and instrument those first.
Why accuracy is the hardest dimension
Validity and completeness can be checked by a machine against fixed rules. Accuracy cannot, because a value can be perfectly valid and still wrong. The fix is not a cleverer rule but a trusted reference to compare against. This is where good data provenance pays off: if you know which source each field came from and when, you can reconcile conflicts instead of guessing. That is the practical link between quality and data lineage for car dealers - lineage tells you where a value came from, which is what lets you judge whether it is right.
How to measure it without boiling the ocean
A common mistake is to attempt a full audit of every field across every record before doing anything. That stalls. A more useful approach is to define a small set of quality checks, run them on a schedule, and express the results as a handful of metrics a manager can read in seconds.
- Define the rules. Write down what "valid" means for the fields that matter: VIN format and check digit, mileage within a plausible range, price present and positive, status from an allowed list, first-registration date not in the future.
- Score each rule. For every rule, compute the proportion of records that pass. A VIN validity rate expressed as a percentage is something you can track week on week.
- Aggregate, but keep the detail. A single headline score is good for a dashboard, but you must be able to drill into the failing records, because the score only matters if someone can act on the list behind it.
- Set thresholds and alert on drift. Decide what is acceptable, then watch for the number falling. A sudden drop in completeness usually means an upstream feed changed or broke.
Fixing it: validation, deduplication and a remediation loop
Measuring dirty data is only useful if it leads to clean data. Three mechanisms do most of the work.
Validate at the point of entry
The cheapest error to fix is the one that never enters the system. Put validation where data is created or imported: reject a VIN that fails its check digit, refuse a mileage outside a sane band, require the fields a valuation depends on before a record can be saved. Catching errors here is far cheaper than discovering them in a pricing report later. If VIN handling is unfamiliar territory, VIN decoding explained covers why the 17-character structure and check digit make such a reliable validation gate.
Deduplicate on stable keys
Match on identifiers that do not vary: VIN first, then registration. Avoid keying on free-text make, model or trim, because "Volkswagen", "VW" and "vw" are the same manufacturer to a human and three different values to a database. When you find duplicates, decide a survivorship rule - which record wins, and which fields are merged from the loser - before you start deleting, so the merge is repeatable rather than ad hoc.
Close the loop with remediation and monitoring
Cleaning is not a one-off project. Records drift: prices go stale, statuses lag, feeds change shape. A durable approach runs the checks continuously, routes failures to whoever can fix them, and re-checks after the fix. This is the difference between a one-time clean-up and a data-quality discipline. It also depends on data not being trapped in disconnected systems - when the same car lives in three tools that never reconcile, you cannot dedupe or monitor it. That fragmentation is the subject of dealership data silos, and resolving it is a precondition for sustained quality.
Where data quality meets pricing and stock
The reason any of this matters is that vehicle records feed decisions. A valuation model is only as trustworthy as the mileage, specification and condition data it reads. A days-to-sell or margin-at-risk metric is only as reliable as the stock list it runs on. Dirty inputs do not announce themselves in the output - they just make the number quietly wrong, which is the worst failure mode because it looks like a confident answer.
This is also why a clean record set and a single source of truth for vehicle data reinforce each other. A single source of truth is the architecture that gives you one place to clean; data quality is the discipline that keeps what lives there correct. You need both. One master copy full of duplicates and bad VINs is no better than several scattered ones.
Where VehIQ fits
VehIQ is being built so that data quality is a property of the platform rather than a separate clean-up project bolted on afterwards. The design centres on canonical European vehicle data with field-level lineage, so every value is intended to carry a record of where it came from - which is exactly what you need to reconcile conflicts and judge accuracy. Validation, deduplication on stable keys and freshness checks are designed to sit inside the data layer, not in a spreadsheet someone runs occasionally.
VehIQ is pre-seed and early in its build, so this is the intended design rather than a deployed track record. The reason it is worth stating now is that the valuation and inventory intelligence VehIQ is designed to provide - valuations that show their sources and a confidence interval, days-to-sell and margin-at-risk signals - are only as trustworthy as the records beneath them. Clean, well-described vehicle data is not a feature on top of that work. It is the foundation it stands on, in open formats the customer owns.