What is a data quality audit, here
A data quality audit measures how far the collected data is from the events that produced it. Four distances matter for analytics data.
- Completeness. Events that happened and never arrived. The checkout that ran inside a child frame and was invisible to a top-window reading is complete in the warehouse only if the collector saw the frame.
- Uniqueness. Events that arrived more than once. A checkout progress event fired six times in one pass and begin checkout fired twice. Any funnel or revenue model on those rows inherits the duplication.
- Validity. Fields whose values are legal but false. On one national retailer's site, closing the mandatory cookie notice leaves every consent category on, so the consent field is a constant across the history. It passes a null check. It fails the truth check.
- Consistency. The same entity named or keyed differently on different pages. A product key that takes a different form on different templates joins to itself as different products.
The data quality audit process
- Profile the table.Standard checks: nulls, cardinality, ranges, duplicates, distribution over time. This is the part every tool does.
- Measure the site.Real sessions on real devices, per consent state, network recorded, every beacon decoded. This is the part tools do not do, and it is where the constants and the doubles come from. Method on /tracking-audit.
- Diff the two.For each column that matters to a business question: what the site sent, what the table holds, and the distance. Counts derived two ways and published only when they agree.
- Trace each fault to its cause.Collection, taxonomy or consent. A constant consent field is a consent fault. A doubled event is a collection fault. A product key that changes by page is a taxonomy fault.
- Grade the questions.Forty business questions asked of the table, or of the AI that reads it, each graded against the measured truth. That is the answer audit, and it is where a data quality audit earns its keep.
Analytics data quality is what the AI inherits
Amplitude's documentation says the more complete the setup, the more context its agents have to work with. Adobe's Data Insights Agent builds its knowledge base from the component names of a data view rather than from the data points. Shopify documents that a cookie banner reduces session counts and the metrics built on them, including conversion rates.
None of the three assistants can see a fault the table hides. Marketing data quality is therefore not a data team concern any more. It is the ceiling on every answer the assistant gives.
Data quality monitoring after the audit
The diff becomes the monitor. Each column that mattered gets its expected shape from the measured site: this event fires once per step, this field varies by consent choice, this key has one form. After each release, the table is checked against that shape, and the release passes or names the column that moved. That is the re-audit tier on /answer-audit.
What you receive
- A profile of the table, with the standard checks.
- A site measurement per page and per consent state, decoded.
- A column-by-column diff with the distance and the cause.
- A fix list for the container owner and the taxonomy owner.
- The graded question set, if the engagement includes the answer audit.
Related: the data layer, Google Consent Mode v2, how we grade, cases.