Data Quality Basics: Clean Data Before You Analyse

By Grocito13 October 2026

Data quality is how well your data is fit for the decisions you make with it. To improve it, measure accuracy, completeness, consistency, timeliness, uniqueness and validity, find where errors enter, validate data when it is first entered, agree shared definitions and give every important dataset an owner. Cleaning once helps briefly; preventing bad data keeps dashboards and AI trustworthy.

Key takeaways

Data quality has six common dimensions: accuracy, completeness, consistency, timeliness, uniqueness and validity.
Most bad data comes from manual entry, merged systems, integrations and definitions that change quietly.
Profile your data first. A simple audit of a few key tables usually reveals the biggest problems.
Prevention beats repair: validate at the point of entry and monitor pipelines for surprises.
Shared definitions and named owners matter as much as any tool.

The six dimensions of data quality

Quality is not one thing, so name what you are checking.

Accuracy. Does the value match reality? A customer's phone number is accurate only if it reaches them.
Completeness. Are required fields filled in? Missing postcodes or lead sources limit analysis.
Consistency. Does the same fact agree across systems? Revenue should not differ between the CRM and the finance tool without an explanation.
Timeliness. Is the data recent enough for the decision? Last month's stock levels are of little use for today's ordering.
Uniqueness. Is each real-world thing recorded once? Duplicate customers inflate counts and split histories.
Validity. Does the value follow the rules? Dates should be dates, statuses should come from an agreed list and emails should look like emails.

Where bad data comes from

Manual entry

People type quickly, make typos, use different spellings and skip optional fields. Free-text boxes where a dropdown would do are a classic cause.

Merges and migrations

Combining two systems or importing a spreadsheet brings duplicates, mismatched formats and conflicting records.

Integrations

When tools sync with each other, a mapping error can copy the wrong field, overwrite good data or create loops. Our guide on data pipelines for growing companies covers how to build these flows more reliably.

Changing definitions

A sales team starts counting a qualified lead differently, or a status is renamed, and trends silently break. The data is not wrong in any single row, but it is no longer comparable.

Profile first: a simple data audit

Before fixing anything, look at what you have. You do not need special software to begin.

Choose three to five tables that feed important reports, such as customers, orders and leads.
For each field, check how many values are empty, how many distinct values exist and what the unusual values look like.
Look for impossible entries: future birth dates, negative quantities, test records and dummy names.
Count likely duplicates using email, phone or name combinations.
Compare totals between systems for one period and note differences.
Rank the problems by how much they affect decisions, not by how many rows they touch.

Write the findings down. The audit gives you a short, prioritised list instead of an overwhelming cleanup project.

Prevent problems at the point of entry

Fixing data after the fact is slower and more expensive than stopping errors as they are typed.

Use dropdowns, pick lists and date pickers instead of free text where possible.
Mark essential fields as required, and avoid making everything required, because people then enter junk to get past the form.
Validate formats for emails, phone numbers and postcodes.
Check for an existing record before creating a new one.
Standardise capitalisation and spacing automatically.
Give clear error messages that tell people how to fix the entry.

If your forms are custom software, these controls are straightforward to build. If you use off-the-shelf tools, many have validation settings that are left switched off.

Deduplication and master data

Duplicates are among the most damaging problems because they distort counts, split customer histories and cause embarrassing double contacts.

A practical approach:

Decide which fields define a match, such as email or a phone and name combination.
Find candidate duplicates and review them, especially where a wrong merge would be hard to undo.
Choose a surviving record and merge useful fields into it.
Prevent new duplicates with checks at entry.

The broader idea is master data: one agreed, authoritative record for important things such as customers, products and suppliers, which other systems refer to rather than copy. You do not need an expensive platform to start. Even deciding which system is the source of truth for each type of record removes a lot of confusion.

Standard definitions and a data dictionary

Many reporting arguments are really definition arguments. Create a simple data dictionary, a shared document or table with, for each field and metric:

A plain-English definition.
The source system.
Allowed values or format.
The owner.
Any known caveats.

For metrics, write the exact calculation. What counts as an active customer, a closed order or a refunded sale? Our article on business intelligence dashboards and choosing the right metrics explains why agreeing on these before building charts saves rework.

Ownership and stewardship

Data without an owner decays. Assign a named person, the data steward, for each important dataset. Their job is not to type everything in; it is to keep the definitions current, answer questions, approve changes and decide what happens when problems are found.

Keep it light. One owner per major dataset, a short review rhythm and a clear place to report issues is better than a committee that never meets.

Monitoring and alerts in pipelines

Quality is not a one-off project, because new errors arrive every day. Add automatic checks where data moves.

Row counts that suddenly drop or spike.
Required fields that arrive empty.
Values outside allowed lists or ranges.
Data that has not refreshed on schedule.
Totals that no longer reconcile with the source.

Send alerts to the data owner, and treat repeated alerts as a reason to fix the cause rather than silence the notice.

Fix vs prevent

Both are needed, in a sensible order. Fix the problems that affect decisions now, such as duplicates in your main customer list. Then trace each type of error back to its source and add a control there. If you keep cleaning the same issue every month, the real problem is upstream.

Why it matters for dashboards and AI

Dashboards present whatever they are given, and a confident chart built on messy data misleads. People who spot one wrong number tend to distrust all of them, and adoption suffers.

AI systems depend on data quality even more. A model or assistant that reads inconsistent, outdated or duplicated records will produce inconsistent, outdated or duplicated answers, just in fluent language. Cleaning and organising data is often the most valuable preparation for any AI project.

A worked example

Imagine a growing online retailer whose monthly report shows more customers than the finance team expects. A short audit finds that the same shopper appears several times because guest checkouts and account sign-ups create separate records, and that country is typed in several ways. The team agrees that email is the matching field, merges duplicates carefully, switches country to a dropdown, adds a check for existing emails at sign-up and writes a one-page dictionary defining active customer. A daily check alerts the owner if the order feed arrives empty. The next report agrees with finance, and the debate moves from whose number is right to what the number means.

FAQ

What is data quality in simple terms?

It is how reliable and usable your data is for its purpose. Good quality data is accurate, complete, consistent, up to date, unique and valid.

How do I clean data before analysis?

Profile it first, then fix the highest-impact issues such as duplicates, missing values and inconsistent formats. After that, add validation at entry so the same problems do not return.

Who is responsible for data quality?

Everyone who enters or uses data shares some responsibility, but each key dataset should have a named owner who maintains definitions and decides how problems are handled.

Do I need special tools to improve data quality?

Not to start. A spreadsheet audit, entry validation and a shared data dictionary go a long way. Tools become valuable as data volumes and the number of systems grow.

Next steps

Pick the one report your team trusts least, trace each number back to its source and list what is wrong. Fix the biggest issue and add one prevention control. If you would like help with audits, pipelines or reporting, see our data analytics services or contact us.