Data Quality: The Silent Errors Skewing Your Analytics

Conversations about data quality tend to start in a Monday morning meeting. The marketing dashboard says August closed with 412 leads. The CRM says 583. Nobody knows which number is right, and a meeting that was supposed to be about strategy turns into forensics: who exported what, when, from which system.

The uncomfortable part is that there is almost never one big, visible error behind that scene. There are dozens of small ones: a duplicated contact here, a UTM that quietly stopped tagging there, a source field filled in three different ways. None of them crashes anything. All of them distort the analysis.

Data quality is the degree to which a dataset meets criteria such as accuracy, completeness, consistency, timeliness, uniqueness and validity. In practice, your data has quality when it describes reality precisely enough to support a decision. Duplicates, broken tracking and ambiguous metric definitions are the most common defects, and also the hardest ones to notice.

Why are the most expensive data errors the silent ones?

Big errors trigger alarms. If revenue shows up as zero on the dashboard, someone fixes it the same day. An 8% error triggers nothing: the number looks plausible, the meeting moves on, the decision gets made. That is why the real problem is not the company that knows its data is bad. It is the one that trusts numbers that are slightly wrong.

A worked example. An online store records 1,000 conversions in a month, with $50,000 invested in paid media. But 80 of those conversions are the same purchase counted twice, once by the pixel and once by the conversions API. The real CAC is $50,000 ÷ 920 = $54.35, not the $50.00 in the report. It seems minor, until you remember this is the number that sets your break-even ROAS and the bid ceiling for every campaign.

What are the dimensions of data quality?

Data governance literature organizes data quality into six dimensions. More useful than memorizing the names is knowing which question each one answers:

Dimension	Question it answers	Example of failure
Accuracy	Does the record describe the real fact?	A $197 sale recorded as $1,970
Completeness	Are the fields that matter filled in?	40% of leads with no campaign source
Consistency	Is the same data identical across systems?	"NY", "New York" and "new york" in the same column
Timeliness	Does the data reflect the current state?	Inventory last synced 6 days ago
Uniqueness	Does each entity appear only once?	The same customer with 3 CRM records
Validity	Does the data respect the expected format?	A phone field filled with "doesn't have one"

No operation scores well on all six all the time, and that is fine. The right question is not "is my data perfect?" but "in which dimension is the error that changes my decisions?".

The four most common silent errors

1. Duplicates in the CRM

Duplicates are born from integrations, not carelessness. The checkout creates a contact on every purchase, the webinar form creates another one, and the sales team registers a third with the phone number written differently. The result: a base of 10,000 contacts that actually holds 8,700 people.

That inflates every metric calculated "per customer". LTV drops artificially, repeat purchase rate vanishes because purchases get split across records of the same person, and the email flow treats a loyal customer as two strangers.

2. Broken tracking

This is the classic silent error because it fails gradually. A website update removes the pixel from one page, a redirect drops campaign parameters, someone launches ads with utm_source=Facebook while the rest of the team uses lowercase "facebook". Nothing breaks visibly. What happens is worse: "direct" traffic grows and campaigns look less profitable than they are.

Anyone investing in paid media feels this directly in conversion attribution: the channel that lost its tracking becomes an involuntary donor of results to the others. A documented UTM standard and an alert for abnormal drops in events solve a large part of the problem.

3. Inconsistent fields across systems

A distributor wants to know how much it sells per state. In the ERP, the field is a two-letter code. In the e-commerce platform, the state is written out in full. In the sales team's spreadsheet, each rep writes it however they like. Sales by state now depends on who built the report, and each department reaches a different ranking from the same data.

The same goes for dates (is 03/04 March 4th or April 3rd?), for currencies and, above all, for business definitions: if marketing counts a "lead" from the form fill and sales counts it from the first meeting, both are right and the report is wrong.

4. Outdated data scattered across silos

When systems do not talk to each other, someone exports a spreadsheet to "solve it quickly". Three months later, that export has become the official source for a budget decision, and nobody remembers how old it is. Data silos are not just an access problem: they are a versioning problem. Every copy ages at a different pace.

How much does poor data quality cost?

The available numbers are striking. Gartner estimates that poor data quality costs companies an average of $12.9 million per year. And a study published in Harvard Business Review found at least one critical error in 47% of newly created records; only 3% of the datasets evaluated met basic quality standards.

For the day-to-day of a small or mid-sized business, the most useful reference is the 1-10-100 rule, formulated by George Labovitz and Yu Sang Chang: preventing a data error costs 1, fixing it later costs 10, and living with it costs 100, in the form of rework, wrong decisions and poorly served customers. It is an illustrative ruler, not a law of physics, but the order of magnitude matches what you see in real operations: validating a form costs hours; rebuilding six months of reports costs weeks.

Data quality audit checklist

You cannot audit everything, and trying is the fastest route to giving up. Start from the metrics that actually drive decisions and work down from there:

1. List the 5 metrics that drive decisions (CAC, ROAS, conversion rate, average order value, churn, whatever applies to your business) and identify which system each one comes from.
2. Compare sources for the same number. GA4, ad managers and the CRM will never match exactly; what matters is knowing the typical gap. A 10% divergence is expected; 40% is a symptom.
3. Measure duplicates: count distinct records by normalized email and phone and compare against the total base.
4. Audit the last 90 days of UTMs: list the distinct values of utm_source and utm_medium. If "facebook" shows up spelled four ways, you just found budget being misattributed.
5. Measure completeness of critical fields: the percentage of leads with no source, no phone, no deal value.
6. Check the freshness of integrations: when each connection last synced and who gets notified when it fails.
7. Write a metrics dictionary: a single definition of lead, conversion and active customer, with a named owner for each.
8. Automate anomaly alerts: a sharp drop in events is almost never the market; it is usually broken tracking.

On frequency: a full audit works well once a quarter. Tracking and integrations deserve continuous monitoring, because they break without warning.

Frequently asked questions

What is data quality?

It is the degree to which a company's data is accurate, complete, consistent, timely, unique and valid. Quality data describes reality well enough to support decisions without someone having to double-check everything by hand before trusting it.

What are the 6 dimensions of data quality?

Accuracy, completeness, consistency, timeliness, uniqueness and validity. They work as a diagnostic script: instead of saying "the data is bad", you identify whether the problem is a wrong record, an empty field, a divergent format, stale information or a duplicate.

How do you measure data quality in practice?

Pick the metrics that drive decisions and build simple indicators on top of them: duplicate rate, percentage of empty critical fields, divergence between systems for the same number, and the age of the last sync. Measuring a few things consistently beats one giant diagnosis done once.

Is data quality related to privacy laws like GDPR and LGPD?

Yes. Laws like the GDPR and Brazil's LGPD require personal data to be accurate, up to date and kept only as long as necessary, which is in practice a quality requirement. Duplicated, outdated bases increase both regulatory risk and the cost of handling deletion requests.

How often should I audit my data?

A structured audit once a quarter is usually enough for small and mid-sized companies. Integrations and tracking call for continuous monitoring with alerts, because they are the parts that most often break silently between one audit and the next.

Trust is built before the dashboard

The realistic goal is not perfect data; it is data reliable enough to decide on without weekly forensics. That is built with collection standards, monitored integrations and a single definition for each metric, long before any beautiful chart.

It also gets easier when verification stops being a project and becomes a question. With Sherlok, you connect Meta Ads, Google Ads, GA4, your CRM and spreadsheets in one place and simply ask: "how many duplicate contacts do we have?" or "which campaign is arriving without UTMs?". The answer comes ready, and Monday's meeting goes back to being about strategy.