Join GA4's campaign attribution with Robin Signals' consent-independent event stream in BigQuery, using an identity map built from consenting traffic.
Pair BigQuery's scale with Python's statistical libraries: stream events out via the Storage API, model them in Python, and write scores back to the warehouse.
In the AI era, clean behavioural telemetry is no longer just for reports — it is the training data and context behind your models, and your competitive moat.
Tracking has grown from campaign support into core data infrastructure that serves product, data science and compliance teams, and it needs engineering rigour.
Build, score and retrain a conversion propensity model entirely in SQL with BigQuery ML — no Python and no separate ML infrastructure.
The GA4 BigQuery export is nearly unusable without transformation. Here's the dbt model you need on day one.
Safari's ITP caps JavaScript-set cookies at 7 days. Here's exactly how to fix it with server-side GTM.
Step-by-step: version-control your sGTM infrastructure with GitHub, Cloud Build, and Cloud Run.
A 20% traffic drop that falls within the seasonal pattern is expected. Only the residual component represents genuine anomalies.
Multi-touch attribution doesn't have to be a black box. Markov chains derive channel weights from real journey data, an auditable basis for spend decisions.
Client-side event data is fundamentally unreliable. Server-side enrichment and validation transform raw events into defensible data.
The probability that a variant is better, and the expected loss of choosing it, is more actionable for business decisions than a p-value.
Raw event data is an implementation detail. Analysts should query metrics, not event schemas.
When 35% of users decline cookies, the missing data comes from a systematically different population — a bias that distorts standard analytics metrics.
Hand-crafted segments are a useful starting point, but propensity models surface complex interaction effects that are difficult to manually encode.
Avoid overusing GROUP BY for behavioural analysis. Window functions compute metrics without collapsing the event grain.
Correlation is everywhere in analytics data; causation is rare. Quasi-experimental methods separate a redesign's real effect from everything else that changed.
Retention cohort tables treat every non-returning user the same, which is statistically wrong. Kaplan-Meier and Cox models handle that uncertainty honestly.
Rule-based attribution assumes user journeys are sequential. They're not — and treating them that way is a persistent mistake in marketing analytics.
GA4's 30-minute session timeout is a vendor default, not an analytical truth. Make sessionisation a deliberate choice that reflects how people use your product.