Data Quality: Dimensions, Metrics, and Where Marketing Data Fails

Data quality is the degree to which data is fit for the decision being made with it, measured on dimensions such as completeness, accuracy, consistency, timeliness, validity and uniqueness. The published counts differ (five elements, six pillars, seven dimensions) because different frameworks split the same properties differently; the underlying list is stable.

For marketing data, the dominant failure is not on that list. A campaign record can be complete, accurate and timely inside every platform that holds it, and still be useless, because the same campaign was named five different ways by five different teams and nothing joins them. That is a consistency-across-systems failure, and it originates at data entry, not in the pipeline.

What is data quality?

Fitness for the decision being made with it.

That qualifier matters more than it looks. Quality is not an absolute property of a dataset; it is a relationship between the data and a use. Customer records with 5% missing postcodes are high quality for a churn model that never reads postcode and unfit for a logistics plan that does. Asking “is our data good?” has no answer. Asking “is it good enough to decide this?” does.

Most vendor definitions agree on this framing. Informatica describes data quality as the measure of how well-suited a dataset is to its specific purpose (Informatica, “What is Data Quality?”, accessed 2026-09-11).

The practical consequence is that quality work starts with a decision, not a dataset. If nobody can name the decision a field supports, there is no standard against which to judge it, and any cleanup effort will be arbitrary.

This also explains why “data quality” and “data accuracy” are not synonyms, though they are used as though they were. Accuracy is one property among several, and a dataset can be perfectly accurate and still unfit: last quarter’s figures are accurate and useless for this week’s pacing decision. Fitness is the parent concept; the dimensions below are the ways a dataset can fail it.

The six dimensions, and why the counts differ

Five elements, six pillars and seven dimensions describe the same properties at different granularities.

Search this topic and you will be told there are five elements, six pillars, or seven dimensions, by sources that are all reputable. The disagreement is not substantive. Frameworks differ in how finely they split the same underlying properties, and in whether they treat a property as a dimension or as a consequence of others.

IBM’s treatment is the most widely cited and covers accuracy, completeness, consistency, timeliness, validity and uniqueness (IBM, “What Is Data Quality?”, accessed 2026-09-11). Semarchy publishes an overlapping but differently-scoped set, which is a useful illustration that the boundary is editorial rather than technical (Semarchy, “What is Data Quality? Dimensions, Benefits & Best Practices”, accessed 2026-09-11).

Relevance deserves a note because it is the one most often dropped, and the only one that can be improved by deleting something. A field nobody uses still has to be populated, validated and migrated. Retiring it is a quality gain that no completeness metric will ever show, and reviewing for it periodically is the cheapest quality work available.

How to measure it

A quality metric expresses one dimension as a rate over a defined population.

An impression is not a metric. “Our campaign data is messy” cannot be tracked, targeted or shown to have improved. A metric needs three parts: the dimension, the population, and the rule that decides pass or fail.

Worked example, on the consistency dimension:

• Dimension: consistency of channel values.
• Population: all campaign records created in the last 30 days.
• Rule: the value appears in the approved channel list, exactly.
• Metric: conforming records ÷ total records = conformance rate.

That produces a number you can put a target on. The same template gives you completeness (required fields populated ÷ records), uniqueness (distinct entities ÷ records), validity (format-conforming ÷ records) and timeliness (records arriving within SLA ÷ records).

Two rules keep these honest. Define the population before you measure, or the denominator will quietly shift to flatter the result. And measure at a fixed cadence, because a single reading tells you nothing about whether the problem is growing.

Assurance vs control vs management

Assurance sets expectations, control catches breaches, management owns the loop.

The three terms are used interchangeably and mean different things.

What follows covers the property and how to measure it. The management discipline — who owns the loop, how remediation is prioritized, how the program is run — is a separate topic with its own page.

The management discipline: Data quality management — ownership, remediation and running the program.

Running an assessment

Profile a real dataset against each dimension and record the rate, not the impression.

A first assessment does not need a tool. It needs one dataset, the six dimensions, and somebody willing to write down numbers.

1. Pick one dataset tied to a real decision. Campaign records for the last quarter, not “our marketing data.”
2. Write the rule for each dimension before looking at the data. Which fields are required? What is the allowed set for each? What counts as current?
3. Profile against each rule and record a rate. Distinct-value counts on fields that should be closed sets are the fastest way to surface consistency problems.
4. Sort the failures by the decision they break, not by volume. A 2% failure on a field that feeds spend allocation outranks a 30% failure on a field nobody queries.
5. Trace each top failure to where the value was created. This is the step most assessments skip, and it is the one that determines whether any fix is durable.
6. Re-run on a fixed cadence. The first number is a baseline, not a finding.

Step 3 usually produces the moment of recognition. A channel field expected to hold six values turns out to hold ninety, most of them spelling variants of the six.

Where marketing data actually fails

Correct in every system, irreconcilable across all of them.

This is the failure mode no dimensions list names properly, and it is the dominant one in marketing. Each platform holds a record that passes every check you could run against it in isolation. The campaign name is present, well-formed, recent and unique within that system. It simply does not match what the same campaign is called in the four other systems that also hold it.

Data quality problems blocking analytics, attribution and reporting is the most prevalent problem in our customer conversations, raised across 96 enterprise accounts. It almost never arrives described as a quality issue. It arrives as two dashboards disagreeing.

The reason standard quality programs miss it is structural. Profiling tools examine a dataset. This failure is invisible inside any single dataset and only appears at the join. By the time it appears, the data has been created, spent against and reported on, and the values are historical fact.

It also originates outside the perimeter that most quality programs cover. A large share of marketing data is created by agencies and partners, in platforms your data team does not administer, by people who never see your quality rules. A control that runs in your warehouse cannot reach them.

Carhartt’s team described what data quality looks like when it is working.

“The thing with data integrity is…you want it to be unseen. When data is corrupt the business is aware of it, but when data is correct and has integrity, the business doesn’t necessarily see data, they see the story.” — Carhartt

The standards layer: What data standards are and how to apply them — the agreed values a quality rule checks against.

Standards as the quality mechanism

A standard converts a quality expectation into an allowed set of values at entry.

This is the link between the two topics. A quality rule says channel must be consistent. A data standard says channel may contain exactly these eight values, in this format, and names who may add a ninth. The rule is the expectation; the standard is the expectation made checkable.

That conversion is what moves quality work from detection to prevention. A dimension you can only measure is a dimension you can only report on. A dimension expressed as an allowed set can be enforced by the form itself, which is the only mechanism that works on people outside your organization.

Validating data and enforcing compliance before a campaign goes live is one of the most common jobs enterprise teams bring us, across 39 accounts, and the framing is consistently pre-launch rather than post-hoc. The cost asymmetry is the reason: a rejected value costs somebody thirty seconds, and a wrong value that reaches production costs a quarter of fragmented reporting.

The wider program that assigns ownership to each field and standard is data governance; the record of what each field means and who owns it is the data dictionary. Quality is the property those two exist to protect.

Checks that actually catch things

The useful check runs before the value is saved, not after it is reported on.

Most quality programs are built out of checks that run too late to prevent anything. The ones that change outcomes share a shape.

• Closed-list validation at entry. The field offers the approved values and rejects everything else. This single control eliminates the entire class of spelling-variant failures.
• Format rules by pattern. Case, separators and date formats enforced where the value is typed.
• Required-field gates at creation. A record cannot be saved incomplete, so completeness never becomes a backlog.
• Cross-system reconciliation on a schedule. The only check that catches the failure mode above. Count distinct campaign identifiers per system and compare; divergence is the signal.
• Distinct-value monitoring on closed sets. If a six-value field starts showing twelve values, something upstream has changed. This is the cheapest early warning available.

Two supporting practices make the technical controls hold. Ownership — every field has a named owner who can approve a new permitted value, so the list stays current instead of being worked around. And training at the point of adoption, especially for agencies and new teams, because a control people do not understand gets escalated around rather than complied with.

The financial case for prevention over detection is well established: Gartner has put the average cost of poor data quality at $12.9 million per organization per year (Gartner, “How to Improve Your Data Quality”, accessed 2026-09-11).

Frequently asked questions

Sources

Outbound citations, named and dated:

• IBM, “What Is Data Quality?” (accessed 2026-09-11) — the most widely cited dimensions taxonomy, used here as the baseline the three published counts are reconciled against.
• Semarchy, “What is Data Quality? Dimensions, Benefits & Best Practices” (accessed 2026-09-11) — a differently-scoped dimension set; the disagreement with IBM is the point being made.
• Informatica, “What is Data Quality?” (accessed 2026-09-11) — the fitness-for-purpose definition.
• Gartner, “How to Improve Your Data Quality” (accessed 2026-09-11) — the $12.9M average annual cost figure.