A protocol for comparing marketplace attribution methods on one seller group and one window

Freeze the comparison boundary

Set the population, calendar period, outcome and counting unit before looking across results. Holding these constant helps distinguish differences created by the attribution route from differences caused by unlike records. Keep the boundary available as you work, rather than revising it to accommodate a platform's output.

For this comparison, include UK-registered sellers on Amazon.co.uk, eBay.co.uk and Etsy who have at least 100 monthly orders. Use 1 to 31 August 2026, Europe/London, with accepted orders counted by order. These definitions apply to every method; do not let a route silently introduce a different seller group or period.

The desk research closed on 6 September 2026 and included no live marketplace records or effect estimate. Observation records an event, allocation assigns credit under a rule, and causal evaluation estimates change attributable to an intervention. Set the decision using a measurement framework, and remember that attribution methods answer different questions.

Log each route's rules

Record the product definition and click or view window for every method before comparing totals. A shared calendar period does not mean the platforms credit the same interaction. Keeping the settings beside each result makes those differences visible rather than treating them as small print.

Amazon Attribution is free and requires an Amazon Ads account or Brand Registry access. It reports detail page views, add to cart and purchases separately, using click windows of up to 14 days and a separate view window; confirm both in your Amazon Ads console. Keep these measures and windows separate; do not merge their outputs into one undifferentiated figure.

For eBay Promoted Listings Standard, a sale is credited after a click within 30 days, and the ad rate is charged on that sale. Etsy Ads assigns revenue to a click within its 30-day window. Capture both definitions before placing their results beside another route, since matching dates alone does not align the credited interactions.

Shopify's attribution report defaults to a 30-day lookback and last non-direct click. It also offers last-click, first-click, linear and any-click models. Walmart Connect reports click-window attributed sales, commonly over 30 days, but Walmart sellers are outside the population defined above, so it is excluded from this comparison. GA4 standard properties are free, default to data-driven attribution and provide 30, 60 or 90-day lookbacks for purchase and other non-acquisition key events (acquisition events offer 7 or 30).

The PECR rule on equipment storage and access supplies the privacy context for storing or accessing information on user equipment. Permission and statistical validity remain separate checks: permission to collect information does not show that a resulting estimate is sound.

Put the methods on one window

The window is the first thing that stops a like-for-like comparison. Choose a common click window of 14 days, the longest that every method in the stated population can report, and apply it as follows:

The outcome of this step is a short table in your own workings with four columns: method, window set, window achieved and comparable or not. Report only the methods marked comparable side by side, and keep the others in a separate list with the reason. No results are reported in this article, because no live marketplace records were used. It gives the protocol that produces them.

Name what the output supports

Classify every output by its evidence type before comparing the numbers. Recorded states, declared credit, rule-based allocation, platform attribution, adjusted association and tested effect are not interchangeable, even when their totals appear similar. Preserve the method's assumptions and limitations when you present its result.

During the stated period, Amazon Attribution, eBay Promoted Listings Standard and Etsy Ads report credit according to their own rules. None establishes what those same sellers would have sold without the spend. Seller Hub exports support reconciliation of recorded states, but do not tell you why an event happened.

Shopify assigns credit through a model chosen by the merchant, so selecting another model changes the result. Amazon Marketing Cloud is a clean room in which an advertiser writes SQL over its own Amazon Ads events, often with 14-day lookbacks. Its paths and modelled relationships describe associations, not causes.

GA4 comparisons of segments are observational: matching channel, device and country still leaves unmeasured differences. Google Ads conversion lift tests divide an audience into test and control groups, then compare conversions. Use causal wording only within the boundary actually studied.

Require a counterfactual first

Do not use causal language unless a reviewed design has a credible counterfactual and the analysis was specified before results were examined. An attributed total or an adjusted comparison by itself cannot establish that an intervention caused a difference.

Before analysis, specify eligibility, the allocation or identifying event, actual exposure, comparison, primary outcome and later-state window. Also record non-compliance, contamination, missing outcomes, exclusions, stopping logic and uncertainty. The planning template can help make stopping logic explicit. Accepted orders in a short period do not represent settled contribution or durable customer benefit.

A qualified statistician must assess whether the design and analysis suit this case. The credibility of a comparison depends on its stated boundary and identifying assumptions. Keep exclusions, missing outcomes and the stopping decision open to review; a platform label is not a substitute for scrutinising the design.

Reconcile the amount before interpreting it

Do not describe captured amounts as revenue or marketplace-attributed order value as incremental profit. First reconcile refunds, disputes, settlement, costs and accounting treatment. Without that work, an amount may be assigned a meaning that the underlying records do not support.

The VAT record requirements provide recordkeeping context for VAT-registered businesses, not a transaction-specific tax conclusion. Check the records behind each amount before comparing totals. Attribution does not replace that reconciliation, and a result cannot settle how a transaction should be treated.

Consumer protection, product safety, privacy, data protection, accessibility and security remain separate gates. A promising attributed figure cannot justify misleading price presentation, ignoring tracking choice, unsafe product handling or an inaccessible journey. These checks concern distinct obligations, not outcomes that an attribution result can resolve.

Choose the question, then decide

Select the route by the question it can answer, not by how precise its output looks. eBay Seller Hub exports support state control, while a transparent rule such as Shopify last-click can support administrative allocation. Use either only when that is the decision you need to make.

Amazon Attribution, eBay Promoted Listings Standard and Etsy Ads provide supplier-defined views, with their limitations stated. They describe credit under their own rules, not what would have happened without spend. Do not present a labelled platform result as an estimate of incremental impact.

Use GA4 comparisons and Amazon Marketing Cloud queries to explore patterns, not to claim causal certainty. Reserve causal claims for a reviewed design with defensible comparison evidence, such as a Google Ads conversion lift test. If the population, state or identity cannot be aligned, mark the methods INCOMPARABLE and stop.

Keep any final decision within the defined seller group, order state, Europe/London period and counting unit. If any boundary cannot be matched, do not combine unlike outputs into one result. The defensible conclusion may be that the available records do not permit comparison, rather than that one method has prevailed.