Direct answer: do not treat the platform’s conversion total as one homogeneous time series. Maintain a ledger that identifies whether outcomes were directly observed, recovered through enhanced matching, imported from CRM, modeled, or reconciled to revenue. When the composition changes, pause simple before-and-after claims and explain the delta before reallocating budget. Platform AI may improve attribution; it does not prove that the campaign caused more people to buy.
The symptom appears when conversions and ROAS rise after Consent Mode, enhanced conversions, a new tag, offline imports, or an attribution-window change while orders, opportunities, or cash remain nearly flat. That does not prove vendor error or growth. Legitimate observability recovery, processing lag, improved deduplication, event redefinition, and real demand can occur together.
Create a Conversion Provenance Ledger
- Event — conversion action, definition version, origin, and optimization objective.
- Class — observed, enhanced-matched, imported, modeled, or reconciled to a financial outcome.
- Identity — click ID, session, lead, contact, opportunity, order, and transaction ID; keep ambiguous matches separate.
- Time — interaction, conversion, upload, processing, attribution, revision, and maturity.
- Value — estimated, attributed, contracted, billed, or collected; currency and margin rule.
- Consent and use — permitted data, destination, market, purpose, and policy version.
- Correction — duplicate, cancellation, return, reopened deal, value adjustment, and replay.
- Decision — reporting, bidding, budget, or experiment; not every evidence class serves every decision.
- Evidence — payload, API response, diagnostics, version, and aggregate reconciliation.
The Ledger never tries to identify a modeled person. It preserves the total’s composition and makes a series break explainable. A combined total may be useful for optimization inside a platform and still be insufficient for an incremental-revenue claim.
Use five evidence levels
- Observed — an event linked with available identifiers and valid under the current definition.
- Enhanced — authorized first-party matching increases observability; it remains attribution rather than causality.
- Imported — a CRM or commerce result sent with ID, time, value, and adjustment policy; quality depends on the source system.
- Modeled — an aggregate estimate for unobservable paths; it is neither a person list nor confirmed cash.
- Reconciled — an order, contract, invoice, or collection compared with media events at an agreed level and window.
No level is universally best. Bidding needs a timely signal; finance needs a durable result; a causal question needs a comparator. The error is using one number to answer all three.
Diagnose the increase before moving budget
- Same event and definition? Compare versions and activation dates.
- Same observed/modeled composition? Record the mix before and after.
- Same denominator? Check sessions, clicks, leads, orders, and eligible opportunities.
- Same window and attribution? Wait for documented maturity; GA4 attribution can revise for days.
- Same operating outcome? Reconcile CRM, orders, cancellations, and revenue.
- Same market and mix? Separate country, channel, product, device, and consent.
- Bidding changed too? Preserve an experiment or controlled rollout so measurement and intervention are not conflated.
Example: conversions rise 22%, orders rise 3%
Hypothetical example: after advanced Consent Mode and enhanced conversions go live, an advertiser reports 1,000 conversions instead of 820. The Ledger finds 90 additional enhanced matches, 70 modeled conversions, and 20 previously missing uploads. Unique orders rise from 410 to 422 with stable cancellations. The sound conclusion is that observability improved and a small observed increase exists—not that demand grew 22% or budget should rise 22%.
Choose the action by the question
- Operate bidding — validate event, consent, diagnostics, volume, duplicates, and stability; use supported platform totals with guardrails.
- Explain performance — separate observed, modeled, imported, and revised amounts whenever the interface or export allows it.
- Reallocate budget — combine platform signals with mature CRM or commerce outcomes, marginal cost, capacity, and testing.
- Claim revenue — use the defined financial milestone, not only attributed conversion.
- Estimate incrementality — use an experiment or defensible design; matching and modeling do not create a counterfactual.
Scope a four-week diagnostic
- Week 1 — choose two campaigns, one conversion, and one downstream result; freeze definitions and dates.
- Week 2 — inventory tags, consent, enhanced conversions, uploads, IDs, windows, and diagnostics.
- Week 3 — reconcile samples and totals across platform, analytics, CRM, and orders; classify differences.
- Week 4 — release the Ledger, maturity rule, break alerts, budget policy, and correction backlog.
Cost rises with platforms, accounts, markets, consent systems, events, domains, apps, CRM, imports, currencies, refunds, granularity, and historical reconstruction. A bounded audit can use exports and cohort reconciliation; continuous operation requires APIs, versioned storage, observability, access, maintenance, and owners. Finance defines the economic milestone; marketing owns the decision; data owns reconciliation; privacy validates permitted use; agencies log changes; platforms own their models.
Measure decision quality, not only recovered coverage
- Observed, enhanced, imported, and modeled coverage by event and market.
- Duplicates, adjustments, ambiguous matches, rejected uploads, and delay.
- Gap among attributed conversion, opportunity, order, cancellation, and mature revenue.
- Budget changes justified by evidence available at the time.
- Scenario error, cost per reconciled outcome, and investigation time.
- Incrementality when a comparator exists; explicit attribution language otherwise.
Clarify media’s role with https://makinai.co/insights/en/paid-media-ai-beyond-google-meta-automation, close outcomes through CRM with https://makinai.co/insights/en/ai-crm-revenue-feedback-loop-ad-platforms, respect maturation through https://makinai.co/insights/en/attribution-window-revenue-maturation-ai-marketing-budget, and separate closed-won from revenue with https://makinai.co/insights/en/crm-closed-won-booked-billed-collected-revenue-ai. If the dashboard rose after a measurement change and nobody can explain the composition, bring MAKINAI at https://makinai.co/services/en/digital-marketing-media-performance-growth-agency one conversion, two periods, the change log, and anonymized CRM or order totals. We can scope the smallest diagnostic that protects the next budget decision.
The sources explain how Google models, imports, reports, and revises data; they do not guarantee advertiser results or turn attribution into causality. Interfaces may limit compositional disclosure. Applicable privacy and consent requirements need review. Without a comparator, report attribution and reconciliation—not incrementality.