Why Google Ads, GA4 and your CRM disagree

Google Ads, GA4 and the CRM can disagree about the same customer because they count different things. Start with one record and establish what each system claims. Missing data, overwritten source fields and different credit rules need different repairs.

Decide before you build the dashboard

Reporting disputes need different remedies. A missing sale requires reconciliation. A source overwritten by a later form requires a data-flow fix. Different credit under different models requires an explanation of the rules. Whether a channel creates additional business may require an experiment. Name which problem you have before rebuilding the dashboard. A more elaborate report can preserve the same confusion if the team still expects one number to settle four different questions.

Begin with the event

What counts as a lead, qualified conversation, opportunity or customer? Can the same person create multiple events? Are cancellations, duplicates and spam handled consistently?

Without shared definitions, a reconciliation meeting becomes an argument about numbers that were never intended to match. Write the rules in plain language and make the report follow them.

Explain why systems differ

Platforms may use different windows, interaction types, identity matching and modelling. Analytics sees only the information available to it. The CRM may store a first source, a latest source or a manually selected value.

Do not average the totals to create a more comfortable number. Decide which system is suitable for each question and preserve a record of material tracking changes.

A measurement contract for the team

This contract reduces false precision without forcing the business to stop making decisions until measurement is perfect.

Audit the handoff

Trace a sample of real enquiries from entry page through submission, routing, qualification and opportunity creation. Check timestamps, campaign identifiers and source fields. A visible sample can expose a logic error faster than another dashboard layer.

Keep the process proportionate. You need enough evidence to find the problem, not a months-long data project before fixing a broken form or duplicated event.

Make decisions under uncertainty

When the data is incomplete, state the hypothesis, evidence, risk and next check. A bounded test can be more useful than committing a large budget based on a fragile attribution model.

For larger causal questions, consider an appropriate experimental design with the volume and operational support it requires. Not every business can run a clean holdout, and a poor experiment can create its own false confidence.

A realistic SEO investment model keeps costs and timing visible. For paid acquisition, the incremental scaling calculation helps test whether the next budget increase has earned its place.

Stop making one chart carry every question

Report delivery, journey health and commercial quality separately, then connect them in the decision narrative. That makes disagreements easier to diagnose. It also makes unknowns visible instead of burying them in an “other” category nobody reviews.

One customer, three claims on the revenue

Suppose a hypothetical customer first reads an SEO guide, later clicks a paid ad and finally returns directly to request a demo. Your SEO report, ad platform and last-visit report can describe different parts of that journey. Adding their claimed revenue together would count the customer more than once.

I would reconcile distinct customers and recognised outcomes in the CRM before negotiating channel credit. Record attribution definitions and time windows beside the report. Two models are not contradictory simply because they answer different questions.

For budget decisions, add economics and experiments where feasible. A modelled share of revenue is not proof that removing a channel would remove that revenue. Use bounded changes with guardrails to investigate the decision you actually face.

Keep one reconciled commercial total. Use channel views to investigate decisions, with their missing journeys and attribution rules visible.

Reconcile one customer without rewriting its history

Choose an authorised customer record and list the known discovery, enquiry, opportunity and purchase events. Keep each system’s timestamp and attribution definition beside its claim. Do not collapse them into a single source field yet.

Check whether missing identifiers, overwritten fields or duplicate records explain a disagreement. If the difference comes from attribution rules rather than missing data, preserve both views with clear labels.

Use the CRM or finance record to reconcile the distinct sales result. Then decide which reporting view serves the next budget question. A first-touch view can describe known discovery without proving incremental acquisition; a last-touch view can describe a preceding interaction without erasing earlier influence. The reconciliation should clarify the journey, not rewrite it to award a winner.

Trace records to find the failure, not to win the argument

Select an authorised sample spanning important sources and outcomes. Follow each record from entry and submission into routing, qualification and later stages. Check whether identifiers, source fields and timestamps survive the journey as intended.

Keep the review focused on the information needed. Do not export unnecessary personal data into a public document or an unapproved tool. The goal is to understand the flow, and a responsibly prepared diagnostic sheet can usually do that without exposing the full record.

Classify discrepancies. Some may be expected differences in definitions. Others may reveal a real implementation problem. This distinction prevents the team from spending weeks trying to make two fundamentally different systems produce identical totals.

Decide when an experiment is worth the effort

Attribution allocates credit under a model. A causal question asks whether an intervention created an outcome that would not otherwise have occurred. For an important budget decision, that may justify a suitable experiment if the business has enough volume and operational support.

Define the intervention, comparison, outcome and risks before launch. Consider contamination, timing and the available sample. An experiment that cannot meaningfully distinguish the expected effect may consume budget without resolving the question.

If a strong experiment is not feasible, use the best available evidence and make a bounded decision. State the limitation. Do not relabel a before-and-after comparison as an incrementality test because the name sounds more rigorous.

Preserve self-reported discovery as a separate signal

Prospects can tell you how they encountered the business, but the answer is imperfect and often describes one part of a longer journey. Store it separately from captured source data. That lets the team compare signals without overwriting one with the other.

Review the answers qualitatively. They may reveal a professional community, an AI answer or a referral path not visible in analytics. Do not use them to assign exact causal credit to every sale. They are valuable information with a different collection method.

Use uncertainty to size the next decision

When evidence is weak, reduce the commitment and define what the next step will teach you. A limited test, a measurement repair or a focused page improvement may be more sensible than an account-wide change. The response to uncertainty should be better decision design, not paralysis or invented precision.

When evidence is stronger, explain what supports the larger commitment and what could overturn it. This makes the business case easier to review and creates a reasoned stopping rule if conditions change.

Use GA4 for a focused investigation

In GA4, open Explore and create a Free-form exploration. Add the dimensions and metrics relevant to your question, use rows for the dimension you want to compare and apply a filter or segment to narrow the analysis. Google’s Free-form exploration guide explains that interface.

For a landing-page investigation, compare page groups using consistently defined outcomes. Treat the result as observed website behaviour under your tracking setup. It does not reconstruct every unobserved journey or prove incrementality. If consent gaps or missing events materially affect the comparison, state that before presenting precise percentages to leadership.

A useful join before a perfect model

The qualified-demo acquisition case connects campaign information with CRM qualification. That join answers a practical question: which acquisition groups produce useful demos? It does not reconstruct every earlier influence on the buyer. This is the level of measurement to establish first when the business cannot reliably connect a paid lead with its sales outcome. Broader attribution can add context once the underlying records are dependable.

A reconciled report can still answer the wrong budget question

You can reconcile every recorded sale and still not know how much business a channel created beyond what would have happened without it. Data completeness and causal evidence are different achievements.

Use the reconciled total to establish what the business sold. Use attribution views to describe known journeys under stated rules. For an incremental budget question, consider an appropriate bounded experiment and its operational constraints.

When a test is not practical, make the allocation with explicit uncertainty, economics and guardrails. Do not turn a precise modelled revenue share into a causal claim. The report earns trust by supporting the decision it can answer and admitting where leadership still has to exercise judgement.

Keep the account separate from the individual contact

One buying contact can submit a form after another colleague researched the business. A CRM join that follows only the submitter may be internally consistent and still omit part of the account journey. Where the available records allow it, retain contact and account identifiers separately and document the joining rules. Avoid assigning every earlier interaction to the final buyer without evidence. The reporting question is whether the view helps the next decision, not whether it creates a complete story for every contract. Unknown interactions should remain unknown.

Making attribution useful enough to act on

Is last-click attribution useless?

No. It can answer a narrow question about the last observed interaction. Problems arise when it is treated as a complete description of influence.

Can we work before attribution is perfect?

Yes. Fix important measurement gaps while making proportionate decisions with the evidence available. Perfect attribution is not a realistic prerequisite for useful progress.

Sources and further reading

Google Analytics: attribution: How reporting models assign credit.

Google Analytics: Free-form exploration: Tables, segments, dimensions and metrics.

Microsoft Research: sample ratio mismatch in A/B tests: Experiment data-quality diagnosis.

References checked 9 October 2026. Sources support the documented facts and tool behaviour; the recommendations and worked examples explain how I would apply them.