Marketing analytics turns observed interactions into evidence about customer behaviour, acquisition and conversion. Attribution assigns credit for recorded outcomes across observed touchpoints. Neither automatically establishes that marketing caused incremental pipeline or revenue. A growth team needs to distinguish measurement validity, descriptive contribution and causal incrementality—then decide which question actually matters before selecting a report or model.
For a B2B SaaS company with a long buying cycle, a visitor may discover a Guide through search, hear a podcast, receive a colleague’s recommendation and later request a demo through a branded query. The platform may attribute the recorded key event to the last measurable touch or distribute credit across observed paths. Both choices leave unobserved exposures and may miss the subsequent account-level sale. Strong analysis is useful precisely because it makes such limitations explicit.
Executive takeaways
- Define the business decision, unit of analysis, event meaning, cohort window and consent scope before creating dashboards.
- Verify real event delivery and CRM record matching; a browser success event is not a qualified lead.
- Separate channel reporting, modeled attribution and experiments for incremental effect.
- Use a hierarchy: instrumentation health → useful engagement → durable receipt → accepted opportunity → mature business outcome.
- Interpret uncertainty by segment and time; never compare an immature acquisition cohort to a mature retention cohort.
1. Begin with the decision, then select the metric
A founder asking “which channel should receive more investment?” needs a different evidence design from a marketing manager asking “did our new form work?” The first may require contribution estimates and tests; the second needs instrumentation parity and actual receipt. Write a question statement with intervention, population, outcome, window and decision threshold. An analytics event without such context is a count, not a conclusion.
Define the object grain: session, user, contact, account, deal or customer. Two people from one organisation and three website visits by one person cannot be compared by simply joining total counts. The analysis should document eligibility, exclusions, bot filtering where available, source identity and known blind spots. Preserve a clear definition of “unknown” rather than assigning untracked activity to the most convenient channel.
Explanatory decision framework
An evidence hierarchy
Different questions need different proof
2. Instrument events with a measurement contract
For each important event, record exact trigger, required parameters, deduplication key, timestamp, destination, privacy handling and acceptance test. A form submission that displays success before the server confirms durable receipt should not be named qualified_lead_received. Name the observable boundary honestly: form_attempt, form_validated, enquiry_received and sales_accepted are separate events if those stages actually exist.
Google Analytics’ events guidance describes collecting interactions and using DebugView and reports to verify events. Its Measurement Protocol validation service warns that a standard submission response does not by itself prove the event was accepted as valid. Test front-end, transport and destination separately. Do not send personal identifiers into analytics parameters without a valid, permitted design.
| Boundary | Example measure | Independent evidence | Frequent error |
|---|---|---|---|
| Interaction | Form attempted | Browser trace | Click counted as submission |
| Validation | Input accepted | Server response | Invalid request counted |
| Receipt | Durable enquiry found | Inbox/CRM record | Silent API loss |
| Acceptance | Reviewed and fit | Sales status history | Automated MQL substituted |
| Opportunity | Defined commercial pursuit | Mature deal record | Account/contact double count |
3. Keep attribution and incrementality separate
Last-click reporting describes a rule assigning credit to the final eligible touch. A data-driven model estimates contributions from observed converting and nonconverting paths according to platform methodology. Google’s attribution documentation explains that its account-specific data-driven model distributes key-event credit using observed paths; model settings and lookback windows affect results. These are model outputs, not an experiment showing what would have happened without a channel.
Incrementality asks a counterfactual question: How many outcomes would not have occurred without this marketing activity? Where the decision merits the cost, use an appropriate holdout, geographic test, randomized account cohort or other carefully designed experiment. Balance sample size, spillover, sales interference and privacy. For low-volume B2B programmes, a statistically powered revenue test may be infeasible; use process evidence, buyer research and bounded decisions without claiming causal certainty.
Explanatory decision framework
One buying journey, several views
A credit rule is not a causal result
4. Build a cohort model for B2B revenue
A cohort must have a defined entry event and an observation window long enough for the outcome. If one campaign launched yesterday and another matured for six months, comparing their opportunity conversion is misleading. Report leading indicators while new cohorts age, and show the fraction not yet eligible for downstream evaluation. When people move between accounts or deals are reopened, define consistent linking and date policies.
Distinguish marketing-influenced from marketing-sourced where your organisation uses both labels. “Influenced” might mean a tracked interaction at any point in a buying journey; “sourced” needs an agreed origin rule. Neither should be used as a claim of causal impact. Finance-approved revenue and sales-owned deal stages provide separate reconciliations; avoid labelling a CRM amount as recognised revenue.
5. A numerical example: denominator discipline
Suppose a hypothetical campaign records 1,000 eligible sessions, 80 browser form-success events, 72 durably received unique enquiries, 30 sales-accepted leads and 9 opportunities after an adequate observation period. The reported event rate is 80/1,000 = 8%; the durable enquiry rate is 72/1,000 = 7.2%; acceptance is 30/72 = 41.7%; opportunity progression is 9/30 = 30%. The outcome is not “9% conversion” and none of these values is a Bargaon or industry benchmark.
The eight-event discrepancy between form success and durable receipt may reflect duplicates or transport failure. The 42 records not accepted raise a different question about audience fit or criteria. A decision to increase budget requires investigating both. If the opportunity cohort is still maturing, even the last rate is provisional. Avoid dividing 9 by impressions and calling it customer conversion.
| Business question | Numerator / denominator | Illustrative value | Caveat |
|---|---|---|---|
| Did browser actions occur? | 80 / 1,000 sessions | 8% | Session grain, not people |
| Were real enquiries received? | 72 / 1,000 sessions | 7.2% | Reconciliation needed |
| Did sales accept the requests? | 30 / 72 records | 41.7% | Must define accepted |
| Did accepted leads progress? | 9 / 30 accepted | 30% | Mature cohort required |
6. Diagnose discrepancies before changing spend
Source disagreement can arise from consent, cookie limits, ad blockers, channel definitions, account matching, delayed CRM sync, timezone boundaries or platform attribution settings. Do not treat any single tool as the unquestioned truth. Trace a sample of known events end to end. Preserve raw source and reattribution fields separately. When tools cannot be reconciled exactly, publish the definition and uncertainty instead of manually editing totals to match.
Use guardrails: accepted-lead mix, unsubscribes, landing-page usability, margin or customer retention where relevant. A channel that creates more key events but lower-quality conversations may be operationally worse. A guide that helps a buying group but rarely generates immediate contact should not automatically be considered unsuccessful.
7. Practical analytic design for SaaS growth
Build three dashboards with shared definitions: data quality (events, missing identifiers, receipt mismatches), journey health (relevant engagement, acceptance and aging), and business outcomes (mature opportunities, customers and retention). Keep report cards short and actionable. Include definition, cohort, last data refresh, caveat and accountable owner with each material number.
Create a decision log: “We increased investment because X; observed Y; alternative explanations include Z; next check on date A.” A good dashboard surfaces reasons to ask better questions, not a single magical blended growth score. If a leader cannot name what will change after reviewing a metric, consider removing it from the executive view.
8. An attribution experiment thought exercise
Imagine a SaaS team running paid promotion to a useful technical Guide. A last-click dashboard shows more branded-search enquiries after launch. This could indicate that the promotion contributed to awareness, or that an unrelated product announcement increased demand. Compare pre-period trends, audience overlap and other marketing activity; where possible, create matched or randomized holdouts. If sample and contamination make a causal estimate unreliable, report a directional learning result rather than an invented ROAS or uplift.
Do not interpret ad-platform conversion values, analytics credit and signed contracts as directly interchangeable. The measurement model and finance model need a documented reconciliation and real dates.
9. A ninety-day measurement plan
Days 1–30 — define: map top decisions, event schema, consent, durable destinations, CRM object grain and cohort windows. Build sample event traces and identify missing data.
Days 31–60 — validate: repair the biggest reconciliation gap, test event delivery and deduplication, compare analytics with actual accepted records and establish a baseline for each eligible segment.
Days 61–90 — decide: run one focused channel or journey experiment where feasible, review uncertainty and guardrails, and record a reallocation or further measurement decision. These are illustrative phases, not a claim that every revenue outcome matures in three months.
10. Construct an evidence hierarchy for investment decisions
Not every decision needs the same causal confidence. Fixing a broken form requires an end-to-end functional test; comparing two channel messages may use a controlled experiment; changing annual acquisition allocation may need multiple mature cohorts and a broader model of marginal costs. Label evidence by what it can actually support. Operational reconciliation proves the record exists; descriptive trends show association; attribution allocates model credit; an adequately designed experiment estimates a counterfactual. None should be presented as interchangeable.
A useful marketing investment memo includes objective, eligible population, baseline period, costs (including creative and sales capacity), expected mechanism, primary outcome, alternative explanations, uncertainty and decision owner. Record whether the aim is more relevant evaluations, lower operating cost, better retention or incremental revenue. Return to the memo at a pre-agreed review date rather than cherry-picking whatever dashboard appears favourable after the campaign.
| Evidence type | Useful for | Not sufficient for |
|---|---|---|
| Event QA and CRM reconciliation | Confirming data delivery | Incremental growth claim |
| Segmented funnel trend | Diagnosing stage movement | Isolating channel causality |
| Attribution model | Describing credit under assumptions | Counterfactual proof |
| Randomised holdout, well designed | Estimating incremental effect | Every future market or segment |
| Qualitative buyer research | Explaining decision mechanism | Population rate by itself |
Decision example: if paid search has strong attributed conversions but many prospects report they already knew the brand, distinguish demand capture from demand creation. Test a carefully bounded reduction or holdout if feasible, monitor total accepted demand and report confidence. When sample size or spillover makes such a test weak, do not invent incrementality; record the unresolved question and make a reversible allocation choice. Analytical maturity is the discipline to say what would change your mind.
11. Frequently asked questions
Is last click wrong?
It is a descriptive credit rule with a limited question, not a causal measure. It can be useful if you understand what it leaves out.
Can AI-search mentions be tied directly to revenue?
Not reliably by default. Citation or mention sampling, referral visits, received enquiries and opportunities operate at different boundaries and may be incompletely observable.
Why do CRM and analytics totals differ?
They count different objects and can have different identity, privacy, validation and date rules. Diagnose those differences instead of forcing totals together.
What if our B2B sample is too small for a controlled test?
Make a proportionate decision using operational evidence, buyer research and descriptive trends. Record uncertainty; do not manufacture statistical precision.
References and further learning
- Google Analytics — events and parameters.
- Google Analytics — validate Measurement Protocol events.
- Google Analytics — attribution models.
- Google Analytics — key events and conversions.
The purpose of analytics is not to claim perfect knowledge. It is to make consequential growth decisions with explicitly bounded evidence and an honest account of what remains unknown.