Data Trends
Why your dashboard is wrong: data quality before visualisation
GOSPELTRADER Research Desk · 28 August 2026 · 7 min read
Quick answer
A dashboard can only be as honest as the table beneath it. Check duplicates, inconsistent categories, silent type coercion, missing-value codes, timezone drift, late-arriving records and changed definitions before you trust a single chart.
When a client tells us a dashboard 'feels wrong', the chart is almost never the problem. Seven faults account for nearly every case we have investigated.
The seven faults
Each of these produces a chart that renders perfectly and reports the wrong number, which is the most dangerous failure mode there is.
- • Duplicate records from repeated imports, inflating every total
- • Inconsistent category spellings splitting one group into three
- • Silent type coercion turning identifiers into rounded numbers
- • Missing values coded as 0, 999 or blank and then averaged
- • Timezone drift shifting events across day boundaries
- • Late-arriving records making the most recent period look weak
- • A definition that changed mid-series without a note
A ten-minute check before you trust anything
Count the rows. Count the distinct keys. If those two differ and they should not, stop. Then list the distinct values of every categorical column and read them — misspellings are visible instantly. Finally, plot the record count per day: gaps and spikes reveal ingestion faults no summary statistic will show.
-- rows vs distinct keys
select count(*) as rows, count(distinct record_id) as keys from source_table;
-- category sanity
select category, count(*) from source_table group by 1 order by 2 desc;
-- ingestion continuity
select date_trunc('day', created_at) as day, count(*) from source_table group by 1 order by 1;Make the check permanent
One-off cleaning decays. Any pipeline we hand over runs these assertions on every load and fails loudly rather than publishing a quietly wrong figure. A dashboard that refuses to update is recoverable; a dashboard that lies is not.
Frequently asked questions
How long should data cleaning take?
On a first engagement, expect cleaning and validation to take longer than the analysis itself. That ratio is normal and is a sign the work is being done properly.