An ecommerce platform has 500 observed stores in January and 500 in June. It is tempting to describe that as a stable customer base. But the same total could conceal substantial movement in both directions. A net count tells you where the balance ended, not how many stores changed their technology during the period.
EcomToolkit recommends an origin-destination table when the question is about observed platform switching. This article explains how to build and read that table without treating detection changes as confirmed merchant churn. All counts below are fictional and illustrate the method; they are not current adoption statistics for named vendors.
Table of Contents
- Choose the observation unit before counting a switch
- Construct a matched panel with explicit states
- Calculate gross movement before net movement
- Keep denominators attached to every percentage
- Reconcile new observations and disappearing sites separately
- Validate the table before turning it into a headline
- The EcomToolkit view
Choose the observation unit before counting a switch
Decide whether a record represents a domain, a storefront, a merchant or a corporate group. These are different units. A merchant can operate several domains, and a domain can contain several commerce-related technologies. A table of domains cannot automatically support claims about the number of businesses won or lost.
The W3Techs methodology illustrates why provider definitions matter: it surveys websites, does not count subdomains as separate websites and excludes redirected domains. Those choices affect comparability with another provider’s list. Read the methodology attached to your actual dataset before combining observations.
Choose two snapshot dates and retain stable identifiers. Normalise the unit according to a documented policy rather than using fuzzy name matching without review. A domain rename, redirect or regional restructuring can otherwise look like one merchant leaving and another arriving even when the underlying business continues.
Define the platform assignment rule too. If the research needs mutually exclusive states, explain how you resolve multiple detections. Otherwise retain a multi-label analysis and avoid forcing it into a single-platform transition table. Our overlapping technology-count guide covers that denominator problem.
Construct a matched panel with explicit states
A matched panel contains the same observable units at both dates. For each unit, record its initial state and final state, along with classification confidence or review flags. This supports an observed transition. It does not by itself establish when the change happened or why the merchant made it.
Keep unknown classifications visible. A failed crawl is not evidence of closure, and an undetected platform is not proof that commerce has stopped. Where possible, distinguish confirmed closure, newly observed sites, unavailable observations and ambiguous technology assignments. The categories should reflect evidence rather than fill a convenient chart.
The simple example below contains 1,000 matched storefront units that are observable and assigned exclusively to one of two fictional platforms at both dates. It deliberately excludes entries, exits and unknowns so the arithmetic is easy to follow. A real report must disclose those exclusions separately.
| Initial platform | Final Cedar | Final Elm | Initial total |
|---|---|---|---|
| Cedar | 450 | 50 | 500 |
| Elm | 50 | 450 | 500 |
| Final total | 500 | 500 | 1,000 |
Read rows as origins and columns as destinations. The diagonal cells contain units with the same observed assignment at both endpoints. The off-diagonal cells contain observed changes. The layout should state that direction clearly; transposing a table without changing the labels reverses the story.
Calculate gross movement before net movement
In this illustration, 50 units move from Cedar to Elm and 50 move from Elm to Cedar. Gross switching across the two platforms is therefore 100 units, or 10% of the 1,000-unit matched panel. Both platforms finish with the same count they started with.
Cedar’s net flow from this bilateral exchange is 50 arrivals minus 50 departures, which equals zero. Elm’s is also zero. Net flow cancels opposite movements; gross flow preserves their scale. A flat share chart can therefore coexist with meaningful underlying switching activity.
The observed endpoint retention share for Cedar is 450 divided by its initial 500, or 90%. The corresponding departure share is 10%. Elm has the same values in this constructed example. These are site-assignment statistics for the defined panel, not contractual customer retention rates or evidence of satisfaction.
Even the diagonal has limits. A site could leave Cedar and return before the second snapshot, remaining on the diagonal despite an intervening change. Two snapshots measure endpoint transitions, not every migration event. More frequent observations can reveal additional movement but still require careful classification and timing rules.
Keep denominators attached to every percentage
Arrival and departure percentages can use different bases. Fifty arrivals divided by the final Cedar count is 10% here. Fifty departures divided by the initial Cedar count is also 10%, but that equality is an accident of the balanced example. It will not generally hold when the platform’s total changes.
A report of destination share among switchers answers another question. Cedar receives 50 of the 100 switching units, or 50%. That does not mean Cedar has half the ecommerce market, nor that half of all merchants chose it during the period. Non-switchers are excluded from this denominator.
W3Techs usage-change reports provide an example of direction-specific explanations for their percentages. Their technology categories and observation rules must be read with the chart; do not assume every reported technology change represents a complete replacement of an ecommerce platform. The fictional table here uses a stricter exclusive assignment by construction.
| Metric in the fictional panel | Calculation | What it describes |
|---|---|---|
| All observed switches | 100 / 1,000 = 10% | Endpoint changes in the matched panel |
| Cedar departure share | 50 / 500 = 10% | Departures relative to initial Cedar units |
| Cedar endpoint retention | 450 / 500 = 90% | Same assignment at both endpoints |
| Cedar share of switcher destinations | 50 / 100 = 50% | Destination among switching units only |
| Cedar net bilateral flow | 50 minus 50 = 0 | Balance of arrivals and departures |
Reconcile new observations and disappearing sites separately
A production dataset is rarely a perfectly closed panel. Additions can come from newly launched stores, expanded crawl coverage or improved detection. Disappearances can come from closures, redirects, access failures or changes in the source’s inclusion rules. These mechanisms should not all be described as platform acquisition or churn.
Publish a coverage reconciliation alongside the transition analysis. Start with the initial observed population, show matched units, unmatched initial units and newly observed final units. Explain which groups enter the transition-rate denominator and which are reported separately. That lets readers assess how much of the apparent story depends on panel selection.
If many difficult-to-observe sites are excluded, the matched panel may favour stable or easily detected stores. Our platform survivorship-bias guide explains why a tidy retained sample can describe a narrower population than the headline suggests. Coverage quality matters alongside sample size.
For a vendor trend discussion, use verified transition samples to generate questions rather than causal conclusions. An observed move does not identify the reason, implementation quality, merchant profitability or total contract value. Those require additional evidence from interviews, business records or a suitably designed study.
Validate the table before turning it into a headline
Reconcile every row total to the initial assignments and every column total to the final assignments. The grand totals should match for a closed panel. Check that a unit appears exactly once and that its identity rule is stable across snapshots. Review duplicate domains and platform aliases before calculating percentages.
Inspect a sample of both off-diagonal and diagonal records. Off-diagonal review can reveal false migration signals; diagonal review can reveal missed changes or ambiguous multi-platform states. Retain the classification rule version so a later detector update does not silently rewrite the historical interpretation.
Write the headline with the population, period and uncertainty intact. A defensible statement might say that 10% of a defined matched panel changed endpoint assignments between two dates. A claim that 10% of all merchants migrated would need evidence that the panel and measurement support that broader inference.
The EcomToolkit view
Platform movement becomes useful when the report preserves direction, identity and coverage. Gross flows reveal activity hidden by net share, while matched-panel definitions limit what that activity can mean. The method should make market research more careful, not turn a technology detector into an unverified customer ledger.
Bring a platform-usage export and its methodology to an EcomToolkit platform statistics review. The first deliverable should be a reconciled population and readable transition table, followed by specific diligence questions that the observed movement can genuinely support.
