Ecommerce conversion attribution methods compared for one merchant decision

Ecommerce conversion attribution methods compared for one merchant decision: whether a checkout page change at an England store raised retained orders. This asks what each route can say about that proposed store change for an adult England customer journey.

Research date: 6 September 2026. Common inputs: legal seller, product, eligible population, Europe/London period, identity rule, customer states, source versions and missingness. Exclusions: vendor claims, invented data, universal thresholds and claimed results. Winner: none.

What to take away

Methods compared side by side

Direct observation: GA4, Matomo and Snowplow

Input is a defined event stream. Output is a count, rate or distribution under the stated rules. It can show placed orders, payment authorisations or retained orders, but it cannot explain why they changed.

Google Analytics 4 uses event data and has a free tier; its data-driven attribution model requires enough conversions. Matomo offers a cloud plan from about €19 a month and can be self-hosted. Snowplow gives a data pipeline you own.

GA4 suits small stores already using Google. Matomo suits privacy-focused teams that want EU hosting. Snowplow suits teams with data engineers.

The definition must specify numerator, denominator, identity, exclusions, late records, duplicates, bots and staff. Observation is useful for monitoring and exception detection. Its main limit is causal ambiguity.

Cross-system reconciliation: Stripe, Shopify and Xero

Input is records that should describe the same commercial journey. Output is matched and unmatched states across store, payment, fulfilment, support and accounting systems.

The government's online-selling page supports the distinction between order and confirmation. GOV.UK returns guidance supports later cancellation and refund states.

Stripe Sigma can query payment states and expose authorised but uncaptured charges. Shopify payouts and order reports show fulfilled then returned cases. Xero and NetSuite hold accounting states that often disagree with store reports. Reconciliation can identify those mismatches. It does not attribute an effect to a page change.

Descriptive association and platform attribution: Triple Whale, Northbeam and GA4

Association compares variables or cohorts without a designed counterfactual. Platform attribution assigns credit using its declared rule and observable data. The output may help describe which tracked path preceded an event.

Triple Whale and Northbeam are ecommerce attribution platforms. They join click, order and ad-spend data, often through Shopify and server-side tracking.

Google Ads uses data-driven attribution by default for many conversion actions. Eligibility typically needs 300 conversions and 3,000 ad interactions in 30 days. Meta Ads Manager applies click and view windows, and its Aggregated Event Measurement limits iOS users to eight conversion events per domain.

Both depend on identity, permission, coverage and rule choices. The ICO's storage and access guidance remains an independent boundary for cookies and similar technologies. Refused or blocked measurement can change the observable population. Neither approach proves causality.

Marketing mix modelling: Robyn, Meridian and PyMC-Marketing

Input is an aggregated time series of activity, commercial outcomes and relevant controls. Output is a modelled relationship under structural, lag and functional-form assumptions. The analyst must document geography, frequency, period, changes in product, price, stock and channel, diagnostics and uncertainty.

Meta Robyn, Google Meridian and PyMC-Marketing are open-source MMM tools. They suit teams with weekly or monthly data and an analyst. A usable series often runs at least 12 to 24 periods.

They can address questions where individual-level joins are unavailable, but correlated activity and changing conditions can weaken interpretation. A model estimate is not an observed order. It also cannot repair incompatible accounting or customer-state definitions.

Controlled experiment: Optimizely, VWO and GrowthBook

Input is a predeclared eligible population, allocation, intervention, comparison, actual exposure, primary outcome, post-order window, missingness plan, contamination assessment, stopping rule and guardrails. Output is an effect estimate with uncertainty if the design and execution support one. Plan at least two weeks and enough sample for 80% power at the minimum detectable effect.

Optimizely, VWO and GrowthBook run randomised tests on pages or features. Google Optimize was retired on 30 September 2023, so old guides may mislead.

HM Treasury's Magenta Book analytical-methods annex describes experimental and quasi-experimental approaches. Its QPIE guidance explains the counterfactual and why monitoring alone does not establish attribution. These are public-sector methods, not store performance standards.

Decision rule

Choose the least ambitious claim supported by the data and design. Use observation for what happened, reconciliation for state disagreement, attribution for a stated credit rule, modelling for an assumption-bound estimate and experiments for a reviewed causal question.

In practice, that can mean GA4 or Matomo for observation, Stripe Sigma or Xero for reconciliation, Google Ads or Triple Whale for rule-based credit, Robyn or Meridian for modelling, and Optimizely or GrowthBook for experiments.

Keep consumer, privacy, accessibility, safety, payment and security stops outside any method score.

Worked example: one England merchant decision

For illustration, an England store with about 20,000 monthly sessions plans a checkout page change. The buyer question is whether the change raised retained orders. The store chooses a controlled experiment in GrowthBook, not a model or platform report.

The record names the eligible population: adult England visitors who reach checkout. Period: four full weeks. Timezone: Europe/London. Identity: first-party cookie plus server-side event. It lists order states: placed, authorised, fulfilled, retained and refunded. Missingness: measured by exposure logs. Assumptions: stable traffic mix. Uncertainty: a 95% interval.

The primary outcome is retained orders per eligible session after the 30-day return window. Guardrails are average order value and refund rate. The chosen claim: the page change raised retained orders per eligible session within the tested population and period. If allocation fails or contamination passes the predeclared 2% limit, the store reports the observation only.

Before you act

Common questions

What inputs does a controlled experiment require?

A predeclared eligible population, allocation, intervention and comparison. Also actual exposure, primary outcome, post-order window, missingness plan, contamination assessment, stopping rule and guardrails. Output is an effect estimate with uncertainty if the design and execution support one. Optimizely, VWO and GrowthBook support this design; Google Optimize closed on 30 September 2023.

Why can monitoring alone not establish attribution?

The Magenta Book QPIE guidance explains the counterfactual and why monitoring alone does not establish attribution. Observation shows placed orders, payment authorisations or retained orders. Its main limit is causal ambiguity.

How should a team choose between these methods?

Match the method to the buyer question. A count needs observation, state disagreement needs reconciliation, a credit rule needs platform attribution, an assumption-bound estimate needs modelling and a reviewed causal question needs an experiment.

Which named tools fit each method?

GA4, Matomo and Snowplow for observation. Stripe Sigma, Shopify payouts and Xero for reconciliation. Google Ads, Meta Ads Manager, Triple Whale and Northbeam for rule-based attribution. Meta Robyn, Google Meridian and PyMC-Marketing for modelling. Optimizely, VWO and GrowthBook for experiments.