Marketing analytics asks how people find a product and become customers. Product analytics asks whether those customers reach an outcome and keep receiving value. The two perspectives should connect, but their metrics need explicit definitions.
A campaign can produce many sign-ups while very few accounts complete onboarding. A product can retain customers well while acquisition remains expensive. A single “conversion” chart obscures the difference. Start by defining the decision each report should support, then decide which events, records and tools it needs.
Separate the questions before choosing tools
| Question | Unit and time window | Evidence to use |
|---|---|---|
| Which campaign brings useful enquiries? | Acquisition channel and an agreed attribution window | Campaign data connected to qualified enquiries in the CRM |
| Where does onboarding fail? | User or account, across the full onboarding period | Completed workflow steps and visible failures |
| Are customers receiving recurring value? | Account cohorts over a defined interval | Repeated business outcomes, alongside usage |
| Did a product change improve results? | Comparable eligible populations before/after or in an experiment | Defined exposure, outcome and relevant confounding factors |
These distinctions do not require separate software for every report. They require consistent definitions. Acquisition source can be useful in a retention analysis when it is an intentional segment, rather than an unexplained difference between the populations being compared.
What GA4 can and cannot establish
GA4 supports event-based measurement, funnels, path analysis, cohort exploration and User-ID. It can answer useful product questions when the implementation and data model support them. “GA4 cannot follow behaviour over time” is too broad a claim.
The practical question is whether the setup represents your product. A B2B account may contain several people, subscriptions and workspaces. Browser events alone may not describe account activation, an approved application or a renewed contract. Those states may live in your application, billing system or CRM.
Reporting limits, retention settings, consent-dependent collection and identity gaps also affect what can be analysed. Check the constraints of the report you actually use. A dedicated product analytics tool can provide a better working interface, but changing tools will not repair events that describe the wrong outcome.
Model outcomes rather than button presses
For a request workflow, distinguish request_started, request_submitted, request_accepted and request_completed. A click on Submit is evidence of an attempt. The application needs to acknowledge durable acceptance before an analytics event can describe an accepted request.
For each important event, record:
- its business meaning and the component responsible for emitting it;
- the identifiers needed to connect the journey, within the approved privacy design;
- the conditions under which it fires, including retries and errors;
- the event version and required fields;
- an owner who can investigate missing or duplicate data.
Illustrative event names are not a universal taxonomy. Use the vocabulary of the actual product. Keep confidential messages, documents, credentials and unnecessary personal information out of analytics payloads.
Connect people, accounts and commercial records carefully
A browser identifier is not the same as a customer account. The same person can use several devices, and several people can belong to one company. Decide whether a particular report counts browsers, signed-in people, workspaces or paying organisations.
GA4 User-ID can support measurement for signed-in users. It does not automatically supply your account hierarchy or make an anonymous visit identifiable. Document the meaning and limits of the identifiers you use; do not assume that every touchpoint can be linked.
Use application and billing records to reconcile material commercial outcomes. Analytics can explain patterns around a subscription; it should not replace the record that establishes whether that subscription exists or has been paid.
Investigate conflicting numbers before changing the roadmap
Suppose analytics reports 30 completed submissions, but the application contains 12 accepted requests. Before concluding that the interface converts well, check whether the event fires on a click, on a successful response, or repeatedly after a rerender. Check test environments, duplicate retries and spam classification as well.
The example is hypothetical. The investigation is practical:
- Compare the same time window, timezone, hostname and event definition.
- Reconcile a small sample with authoritative application or CRM records.
- Check consent behaviour, blocked requests and identity gaps.
- Check duplicate emission, retries and changes in the event schema.
- Separate test traffic and supplier enquiries from qualified customer outcomes.
- Write down what remains unobserved instead of presenting an exact attribution story.
A collection change can look like a change in demand. Annotate releases affecting consent, routing, forms or event definitions before comparing periods.
Share definitions without building an unnecessary data platform
Small teams can begin with a concise measurement plan, a reliable application database, an analytics tool and a CRM. A warehouse becomes useful when several sources need recurring reconciliation, reproducible historical queries or a shared reporting layer.
GA4's BigQuery export can be part of that architecture. Google documents expected differences between the export and reporting surfaces, including differences in modeling and attribution. Exported data is not automatically an exact copy of every UI total, and warehouse storage does not restore events that were never collected.
Choose technology after defining query patterns, freshness, volume, access and operating capacity. Our ClickHouse and BigQuery comparison covers that infrastructure decision separately.
Turn a report into a defensible decision
Every decision report should state the population, time window, metric definition, source, important exclusions and known collection gaps. Show counts alongside percentages, particularly for small cohorts.
A correlation between feature usage and retention is a reason to investigate, not proof that the feature caused retention. More experienced customers may both use that feature and retain better. Compare appropriate groups, use experiments where feasible and combine behavioural data with customer research.
The useful outcome is a next action with a stated level of confidence: repair a broken step, interview a particular cohort, test a change or gather more evidence. A larger dashboard is not automatically a better decision system.
Where engineering support helps
We help teams define events, reconcile business records and build internal reporting tools. For a website, lead-generation engineering connects the submission flow to the actual sales process. The scope should follow the decision the team cannot currently make reliably.
Common questions
Do we need two analytics tools?
Not necessarily. One tool may serve both perspectives. Separate metric definitions and responsibilities first; add tooling when a specific reporting or operating need justifies it.
Can analytics prove which page produced a lead?
Sometimes a measured journey or campaign parameter provides useful evidence. Consent, device changes, missing identifiers and offline contact can leave gaps. The contact page is a submission location; it is not automatically the first page that influenced the buyer.
Is a backend event always correct?
No. It can still fire twice, describe the wrong state or be lost. Define its meaning, handle retries and reconcile important outcomes with authoritative records.

