What data quality means in marketing
A number has quality when it is complete, consistent, correctly defined and current. Complete: every event that should be recorded is recorded, across every template, device and consent state the plan promises. Consistent: the same thing is called the same name everywhere, so it groups correctly and can be compared over time. Correctly defined: the number computes what its definition in the measurement plan says, with the inclusions and exclusions that definition lists. Current: the data arrives within the latency the decision needs, and nobody is reading last week's number as this week's.
Marketing data has three features that make quality harder than it is in, say, a finance ledger. It is collected by instruments — tags, pixels, scripts — that run on other people's browsers and can be blocked, delayed or broken by a site change nobody told the analyst about. It is collected by several systems with their own definitions, so "a conversion" is several different things. And it is used for decisions quickly, often the same week, so a broken number does damage before anyone checks it.
The practical consequence is that quality is a routine, not a state. The measurement plan defines what correct means; the tracking plan specifies how it is produced; this guide covers how you find out, continuously, whether production matches definition. Where a KPIA metric tied directly to an objective. is reported without that routine behind it, it is an estimate wearing the costume of a measurement.
The failure catalogue
Most problems are one of the following. Learn the symptoms; the causes repeat.
| Problem | Symptom | Usual cause |
|---|---|---|
| Double counting | Conversion rate roughly doubles; conversions exceed orders in the order system | Tag fires on load and on action; two tags for one event; thank-you page reloaded or revisited counts again |
| Missing events | A metric drops to zero or near zero on a date; a template reports nothing | Release changed a selector or data layer key; tag paused; trigger condition no longer met |
| Consent gaps mistaken for errors | Users and sessions fall in some markets; conversions fall more than sessions | Consent banner change; consent mode misconfigured; a tag now correctly waiting for consent it did not wait for before |
| Campaign fragmentation | One campaign under several names; unassigned channel traffic grows | Hand-typed UTMURL tags identifying traffic source. values, mixed case, mediums outside the list, redirects stripping parameters |
| Internal and test traffic | Spikes from one location; events with test names; conversions from staff | Internal traffic filter missing or inactive; test events not stripped; staging site sending to the production property |
| Self-referrals and cross-domain breaks | Own domain appears as a referral source; sessions split at checkout or payment | Cross-domain measurement not configured; a payment provider or subdomain starting a new session |
| Wrong parameter types | Revenue sums to nonsense; a numeric dimension appears as text | Currency symbols or thousands separators in values; a string where a number is expected |
| Definition drift | A metric changes level on a date without a change in the business | Platform redefined a metric; a key event or conversion definition edited; attribution setting changed |
| Report filters and segments | Numbers differ between two reports of the same thing | A filter or comparison left applied; a date range that excludes the last day; a sampled exploration against an unsampled report |
| Cardinality overflow | An "(other)" row appears and grows; a dimension loses detail | Too many distinct values for the documented limit; free-text parameters; unique ids as dimensions |
| Bot and spam traffic | Sessions with near-zero engagement from unexpected places | Known bots filtered by default but not all; referral spam; load testing |
| Latency misread | Yesterday looks terrible every morning | Reading a day before processing has finished; platform conversions still attributing |
The GA4 common mistakes page covers the GA4-specific versions of several of these in troubleshooting form; the table above is tool-agnostic and includes ad platforms and the CRM.
A weekly check routine
The weekly check exists to catch the expensive failures within days. It is a fixed list, run by a named person at a fixed time, with results written down even when everything passes. The list:
- Key events against the source of truth. For each conversion the plan names, compare the analytics count for the last complete week with the order system or CRM count. A difference outside the tolerance recorded in the measurement plan is a finding. Compare the ratio over time, not just the counts: a stable ratio that suddenly moves is the signal.
- Event volume by day. Chart the main events for the last fortnight by day and look for a step change — up for double counting, down for a broken trigger — on a date that matches a release.
- Channel and campaign hygiene. Sort the campaign report for the week by value and scan for case variants, unparseable names and mediums outside the list. Check the unassigned channel's share against its normal level.
- Referral sources. Look for the site's own domain, payment providers and known subdomains in the referral report.
- Internal and test traffic. Confirm the internal filter is active and look for events with test prefixes or a spike from the office location.
- Platform settings diff. Open each ad platform's conversion settings and the analytics key event list and compare with last week's screenshot or export. Changes made by other people are the commonest cause of definition drift.
- Container version. Confirm the live container version matches the one the tracking plan says is current, and read the description of any new version.
Record the date, who ran it, each check's result and any ticket raised. The log is what lets the next audit say "this has been clean since" rather than "we think it is fine". When a step change traces to one affiliate partner rather than to a release, the tutorial on investigating suspected affiliate fraud takes the case from there. Clean sources also make it possible to reconcile figures from queries an AI tool wrote, which the checking AI-written SQL tutorial depends on.
Where reports read from warehouse tables, the mechanical checks in this list, such as volume by day and counts against the source of truth, can also run on every load, and the tutorial on writing data tests for marketing tables turns them into SQL queries that fail loudly.
A weekly check that caught a broken form in four days
An insurance comparison site's analyst runs the check every Monday. On week 14 the ratio of generate_lead
events to CRM leads, normally between 0.97 and 1.03, read 0.61 for the previous week. The event-by-day chart
showed the drop beginning on the Wednesday, the day a release had moved the quote form into a modal. The
container version was unchanged, so the trigger — a visibility rule on the form's success element — had
simply stopped matching. The ticket was raised at 09:40, the trigger was repointed in preview by 11:00, and
the Thursday-to-Sunday gap of roughly 190 missing leads was documented as a known hole rather than a real
decline. The paid search team, who had already drafted a bid reduction based on the apparent fall in lead
rate, cancelled it before it went live.
Running a data quality audit
An audit is the deeper version: everything the weekly check assumes is examined. Run it quarterly, after a change of ownership, before a major reporting change, and whenever the weekly check keeps finding things whose cause is unclear. The method:
Scope from the measurement plan. List every metric the plan reports, its definition, its source of truth and the events behind it. The audit judges each metric against its own definition, not against an idea of what the tool should do.
Inventory the instruments. Export the tag container: every tag, trigger and variable. Export the analytics property's event list, key events, custom definitions, filters and data streams. Export each ad platform's conversion actions and their settings. Map every item to the tracking plan; unmapped items are findings.
Test the critical paths. For each conversion, walk the journey on each major template and device with Tag Manager (GTMTag management system.) preview and the analytics debug view open, and confirm the event fires once with correct parameters. The debugging and preview tutorial covers the mechanics. Include the negative cases: reloads, back navigation, repeat visits.
Reconcile against the source of truth. For the last complete month, compare each conversion count and revenue figure with the order system or CRM, by day and by channel where the source allows. Note the ratio and whether it is stable.
Check configuration against the plan. Cross-domainConfiguring analytics so a visit that moves between a business's own domains stays one session. lists, internal trafficVisits from the business's own offices, staff, agencies and testers, labelled by an address rule and filtered so they do not count as customers. rules, referral exclusions, session timeout, attribution settings, data retention, currency and time zone. Each is a setting someone chose once; confirm it still matches the plan's decision.
Scan the data for the catalogue. Run each failure in the table above as a query or a report view and record what you find.
Write findings as tickets. Each with the metric affected, the period affected, the cause, the fix, the owner and how to verify it. Order by the decisions the affected metric informs, not by how easy the fix is. Reporting and dashboards that consume the data — covered in the reporting and BI hub — are updated last, once the sources are clean.
Fixing root causes
A data quality problem has a symptom in a report and a cause somewhere upstream, and fixing the symptom is how teams end up with a dashboard of adjustments nobody can explain. The causes cluster in four places. Tracing a specific problem back to one of them is the job of the five whys method, which asks why at each step and checks every answer against a log, a tag test or a change record.
The tag container. Triggers that match too much or too little; tags duplicated across workspaces; variables reading the wrong data layer key. The fix is in the container and in the tracking plan's trigger column, which should have been precise enough to prevent it; update both.
The site code. Data layer pushes with wrong types, missing keys on some templates, success states that fire before the server confirms. The fix is a developer ticket written against the tracking plan's data layer specification, with the negative tests added to the release checklist.
The convention. Fragmented campaigns and unassigned traffic are never fixed in the report; they are fixed by the link builder, the controlled vocabulary and the monthly exceptions review in the UTM and naming conventions guide. Historical data can be regrouped with a custom channel group or a cleaned export, but the fix is upstream. The same drift reaches CRM fields such as country and job title, and the tutorial on standardising CRM field values with workflows fixes it as records arrive while keeping the raw value beside the clean one.
Platform configuration. Conversion definitions, attribution settings, filters and linking. These change when someone with access changes them; the fix is the change itself plus a record in the measurement plan of who may change what, and the weekly settings diff that notices when they do.
The discipline that ties these together: every finding updates a document. A trigger fix updates the tracking plan; a naming fix updates the convention; a settings fix updates the measurement plan's configuration record. A fix that leaves the documents unchanged will be undone by the next person who reads them. The measurement planning hub shows where each document sits in the sequence. Each change and each fix also belongs in the measurement change log, so that the next person to see a step in the data can find its cause.
When tracking breaks outright rather than drifting, the tracking outage response SOP explains how to confirm the outage, contain the damage, fix the cause and annotate the affected reports.
For a sudden surge of sessions that may be bots, spam or test traffic rather than people, the diagnosis page for a sudden traffic spike reads the spike by hour, location, hostname and engagement.
Common mistakes
- Patching the report. Adjusting a dashboard to hide a known problem instead of fixing the tag. The adjustment outlives the memory of why it exists.
- No tolerance written down. Without an agreed acceptable difference against the source of truth, every weekly check becomes a debate.
- Checking only when something looks wrong. The expensive failures are the ones that look plausible — a conversion rate that is double but stable.
- Treating consent loss as a bug. A fall in users after a consent change may be the tags finally behaving correctly; check the consent configuration before the trigger.
- Auditing the tool instead of the plan. An audit without the measurement plan's definitions judges the data against the auditor's assumptions.
- One person implements, checks and reports. Separate the verifier from the implementer, as the tracking plan requires.
- Letting findings die in a document. A finding without a ticket, an owner and a verification step is a note, not a fix.
