Measurement
Part of Measuring one customer data journey without pretending correlation is cause
Attribution methods for customer data platforms and when to withhold causal claims
Attribution methods for customer data platforms, covering declared, deterministic, modelled, matched and experimental routes and when to withhold causal claims.
No attribution method for customer data platforms is universally best. This comparison fixes one population and one window for every route.
Adobe Customer Journey Analytics, Salesforce Marketing Cloud Intelligence, Amplitude, Google Analytics 4 and Meta Ads Manager are compared here. AppsFlyer, Adjust and Snowplow join them, alongside the identity-only CDPs Twilio Segment, Tealium and mParticle.
Buyer weights and acceptance thresholds stay blank, and no result is reported. Compare methods only for the same decision, eligible population, outcome event and period. Hold the Europe/London clock and the identity rules fixed. The output should say what the method can support, not borrow causal language from a different design.
What to take away
- No attribution method is universally best; compare only for the same decision, population, outcome, period, clock and identity rules.
- Declared-source and deterministic methods answer who received credit under a convention, not incrementality.
- Supplier model credit is not independent causal evidence and should not be added to warehouse outputs.
- Matched and statistical methods adjust only for measured information, so unmeasured confounding may remain.
- If observation is incomplete or a counterfactual cannot be defended, report descriptive or attributed results only.
Common comparison record
For each route, document the required data, its missingness and the deduplication rule. Add the attribution window, the model or rule version and the stated assumptions. The record also carries the output unit, uncertainty, privacy and PECR effects, the correction route and the failure condition.
Attribution methods compared
Output
- No attribution
- Dated event record
- Declared source
- Credit to self-declared route
- Deterministic rule
- Credit under written window
- Supplier model
- Supplier-defined credit
- Matched observational
- Adjusted association
Main assumption
- No attribution
- Sources reconciled as recorded
- Declared source
- Declaration captured and applicable
- Deterministic rule
- Rule and identifiers complete
- Supplier model
- Documentation and version describe output
- Matched observational
- Measured covariates address differences
Causal limit
- No attribution
- Describes occurrence only
- Declared source
- Recall and selection may affect
- Deterministic rule
- Allocation is not incrementality
- Supplier model
- Vendor credit not causal evidence
- Matched observational
- Unmeasured confounding may remain
The ICO's data minimisation guidance requires attention to whether personal information is adequate, relevant and limited to the stated purpose. It does not select a method for the buyer.
Where credit is computed differs by product. Adobe Customer Journey Analytics applies Attribution IQ models: last touch, first touch, linear, participation and algorithmic attribution. Salesforce Marketing Cloud Intelligence, formerly Datorama, offers rule-based and algorithmic attribution. Amplitude's Attribution chart covers first touch, last touch and even weight.
Google Analytics 4 defaults to data-driven attribution, with a 30-day lookback for acquisition conversions and 90 days for other conversion events. Meta Ads Manager reports on a 7-day click and 1-day view default.
AppsFlyer and Adjust typically count a 7-day click and a 24-hour view on mobile installs. Snowplow's open-source attribution dbt package computes first-touch and last-touch credit in the warehouse.
Output
- No attribution, manual observation
- Dated event record
- Declared source
- Credit to a self-declared route
- Deterministic rule
- Credit under a written window and identity rule
- Supplier or platform model
- Supplier-defined credit
- Matched observational analysis
- Adjusted association
- Statistical association
- Relationship in defined data
- Experiment or quasi-experiment
- Estimated incremental effect
- Adobe Customer Journey Analytics, Attribution IQ
- Credit per configured model inside the CJA dataset
- Salesforce Marketing Cloud Intelligence
- Rule-based or algorithmic credit
- Amplitude Attribution chart
- First touch, last touch or even weight
- Google Analytics 4 data-driven attribution
- Credit on 30-day acquisition and 90-day other-conversion defaults
- Meta Ads Manager
- Credit on a 7-day click, 1-day view default
- AppsFlyer and Adjust
- Install or in-app credit, typically 7-day click, 24-hour view
- Snowplow attribution dbt package
- First-touch or last-touch credit in the warehouse
- Twilio Segment Unify, Tealium AudienceStream, mParticle
- Identity graph and audience output
Main assumption
- No attribution, manual observation
- Sources can be reconciled as recorded
- Declared source
- Declaration is captured and applicable
- Deterministic rule
- Rule and identifiers are complete enough
- Supplier or platform model
- Documentation and model version describe the output
- Matched observational analysis
- Measured covariates address relevant differences
- Statistical association
- Specification and data are suitable
- Experiment or quasi-experiment
- Assignment or identifying assumptions hold
- Adobe Customer Journey Analytics, Attribution IQ
- Identity stitching and the configured lookback hold
- Salesforce Marketing Cloud Intelligence
- Journey inputs cover the stated window
- Amplitude Attribution chart
- Tracked events cover the journey
- Google Analytics 4 data-driven attribution
- Tagging and modelled conversions represent the journey
- Meta Ads Manager
- The pixel or Conversions API captures the event
- AppsFlyer and Adjust
- Device identifiers match accurately enough
- Snowplow attribution dbt package
- Event model and identity stitching are complete
- Twilio Segment Unify, Tealium AudienceStream, mParticle
- Identity resolution resolves the person
Causal limit
- No attribution, manual observation
- Describes occurrence only
- Declared source
- Recall and selection may affect the record
- Deterministic rule
- Allocation is not incrementality
- Supplier or platform model
- Vendor credit is not independent causal evidence
- Matched observational analysis
- Unmeasured confounding may remain
- Statistical association
- Correlation does not establish intervention effect
- Experiment or quasi-experiment
- Result applies only within the design and uncertainty
- Adobe Customer Journey Analytics, Attribution IQ
- Model credit is not incrementality
- Salesforce Marketing Cloud Intelligence
- Vendor credit is not independent causal evidence
- Amplitude Attribution chart
- Allocation is not incrementality
- Google Analytics 4 data-driven attribution
- Platform credit is not incrementality
- Meta Ads Manager
- Platform credit is not incrementality
- AppsFlyer and Adjust
- Credit is not incrementality
- Snowplow attribution dbt package
- Allocation is not incrementality
- Twilio Segment Unify, Tealium AudienceStream, mParticle
- Credit model sits in the destination
Keep identities and windows fixed
Declared-source and deterministic methods can be useful for routing or reporting, but they answer who or what received credit under a convention. A supplier model may use a different identity graph and lookback window from the buyer's warehouse. Do not add their outputs together.
Handling identity uncertainty
Is the person unresolved?
Keep uncertainty visible; do not force join
Proceed with fixed identity and window
Identity spines differ as well. Twilio Segment Unify, Tealium AudienceStream and mParticle each build a profile of the same person under their own resolution rules. Identity uncertainty must remain visible. If an unresolved person is forced into a join, both eligibility and exposure may change.
On its accuracy-principle page, the ICO tells organisations to preserve the source and status of personal information. Its current DUAA review notice is a publication-day recheck trigger.
Ask when a counterfactual is credible
Matched analysis can compare apparently similar records, yet it only adjusts for measured information under the declared model. Statistical association can reveal patterns but does not show what would have happened without the activity.
Experiments use an assignment mechanism to construct a comparison. Quasi-experimental approaches rely on a defensible source of variation and identifying assumptions. Meta Conversion Lift, Google Conversion Lift, geo experiments, Measured and Haus all depend on that assignment holding.
Marketing mix modelling sits outside user-level attribution. Meta's open-source Robyn and Google's Meridian estimate channel effects from aggregate spend and outcome series, so they cannot settle a person-level credit question.
HM Treasury's 2026 Magenta Book evaluation guidance covers experimental, quasi-experimental and theory-based approaches for government evaluation. A qualified evaluator must assess whether any design is credible for this buyer case.
Account for refused or missing observation
Pixels, device identifiers and link tracking may affect which events are observable. On iOS, App Tracking Transparency has gated cross-app identifiers since April 2021, which narrows what AppsFlyer and Adjust can match. Google Analytics 4 consent mode and modelled conversions change what a data-driven model can see.
Dated 29 April 2026, the ICO's final guidance for storage and access technologies is the regulator record checked for this comparison. It does not authorise a configuration. Record refused tracking, unavailable identifiers and suppression as defined states rather than silently discarding them.
Choose the least complex route that answers the decision honestly. If observation is incomplete or a counterfactual cannot be defended, report the result as descriptive or attributed and withhold a causal conclusion.
Before you act
- Fix one decision, population, outcome, period, clock and identity rule.
- Document required data, missingness, deduplication, window, model version and assumptions.
- Keep identity uncertainty visible; do not force unresolved people into joins.
- Record refused tracking, unavailable identifiers and suppression as defined states.
- Do not add supplier model credit to deterministic or declared-source outputs.
- Withhold causal conclusions when observation is incomplete or a counterfactual is indefensible.
Common questions
What should each comparison route document?
Each route should document required data, missingness, deduplication and the attribution window. Record the model or rule version, assumptions, output unit and uncertainty.
Privacy and PECR effects, the correction route and the failure condition belong in the record. ICO data minimisation guidance requires attention to whether personal information is adequate, relevant and limited to the stated purpose.
When can a counterfactual be treated as credible?
Experiments use an assignment mechanism to construct a comparison, while quasi-experimental approaches rely on a defensible source of variation and identifying assumptions. Matched analysis only adjusts for measured information under the declared model, so unmeasured confounding may remain. A qualified evaluator must assess whether any design is credible for this buyer case.
How should refused or missing observation be handled?
Pixels, device identifiers and link tracking may affect which events are observable. Record refused tracking, unavailable identifiers and suppression as defined states rather than silently discarding them. The ICO guidance for storage and access technologies, dated 29 April 2026, is the regulator record checked for this comparison and does not authorise a configuration.
