A purchase after a reminder does not prove the reminder caused it
To decide whether a recovery workflow deserves to continue, counting orders that arrived after a message is not enough. First define which records are eligible, the observation window and the source confirming the order. Then distinguish observed purchases, platform attribution and incremental effect. This guide proposes a way to read results and a fictional example; it presents neither store results nor an automatic Think² measurement function.
Do I want to know what happened after contact?
Link each message and order using identifiers and timestamps within a window defined before the reporting cutoff. This is descriptive tracking, not causal evidence.
Do I want to reproduce a dashboard number?
Retain that report's attribution definition and window. Do not change its denominator or call every attributed order an additional sale.
Do I want to know whether contact produced additional purchases?
You need a comparison designed for that question, ideally randomized before contact among equally eligible records. A small example or before-and-after comparison does not establish the real effect.
Waitlist; immediate access or a store connection is not guaranteed. Spanish destination.
Define the unit before counting
A visitor, cart, checkout, customer and order are different units. Choose one and retain its identifiers. If a customer opens three checkouts and places one order, do not count three recovered purchases. Do not infer a reason for abandonment merely from the absence of an order.
Set eligibility and exclusions
Separate records with insufficient identification, purchases already completed and contacts not authorized or excluded under the applicable rules. Eligibility for analysis does not authorize contact. Record why you exclude each record without exposing unnecessary data.
Fix the cutoff and observation window
Record when the cohort was fixed, when each message was sent and the latest instant for checking orders. Do not silently add an order after the window closes: identify it as an update. If the source is incomplete, leave the outcome unknown rather than turning it into zero.
Count attempts, deliveries and purchases separately
Preparing a message is not sending it; sending does not prove delivery or reading. A reply does not confirm an order or payment. Use only a status you can verify from the source. Do not call an order value collected revenue without confirmed payment.
Do not confuse three readings
An observed purchase after a message is a temporal association in the cohort. A dashboard-attributed purchase satisfies a particular attribution rule. An incremental purchase represents a difference from what would have happened without contact. The first two do not establish the third.
What Shopify documentation calls recovered
The consulted documentation allows the customer to complete checkout after the email either through the link or on their own. That recovered definition does not by itself establish that the message caused the purchase. No store configuration, report available in a particular account or random-testing feature is established here. The following example uses our own rules rather than reproducing a Shopify report.
Fictional example
Fictional data to explain denominators; not a customer sample, performed experiment or Think² results. Dates and counts were chosen for this example.
Cohort fixed on November 2, 2026 at 09:00 UTC; orders observed until November 5 at 09:00 UTC. The fictional dataset includes complete outcome records for that window; completeness is not established for a real store.
Each of the six eligible records represents a different customer with one checkout in this example. No orders are shared between records. These are fictional groups, not a performed randomized assignment.
X1
Purchase completed before fixing the cohort
Exclude from recovery; do not contact again about that checkout.
X2
Contact not authorized for this workflow under the example's rules
Exclude from sending; do not assume authorization.
C1
Eligible, contact group, message sent, order confirmed within the window
One observed purchase; delivery and reading are not assumed.
C2
Eligible, contact group, message sent, order confirmed within the window
One observed purchase; this does not mean the message caused it.
C3
Eligible, contact group, message sent, no order in the complete fictional window
Zero orders at this cutoff, not a permanent lost sale.
C4
Eligible, contact group, message sent, no order in the complete fictional window
Zero orders at this cutoff, not a known reason for abandonment.
H1
Eligible, no-contact group, order confirmed within the window
A purchase without a reminder in this example.
H2
Eligible, no-contact group, no order in the complete fictional window
Zero orders at this cutoff, without predicting future behavior.
- 8 initial records − 2 excluded = 6 eligible.
- Contact group: 2 records with an order / 4 eligible assigned records = 50%.
- No-contact group: 1 record with an order / 2 eligible assigned records = 50%.
- Descriptive difference: 50% − 50% = 0 percentage points.
It would be incorrect to count only the two purchases after the message and call them two additional sales. The no-contact group also buys. These invented sample sizes cannot establish a null effect or estimate the real effect; there was no actual assignment, uncertainty calculation or represented population. The example also does not determine how many orders a dashboard using another attribution rule would count.
If the order source for H2 becomes unavailable, you cannot retain the group's rate as though its outcome were known. Recover the evidence or report an incomplete result; unknown does not mean no purchase.
You can assign a collaborator to review authorized exports: identifiers, dates, exclusions, denominators and differences between a dashboard and a report. In a group, marketing and the person who understands the orders can review definitions. Think² offers compatible models, authorized tools and an optional computer per collaborator; verify access and the result. Store instrumentation, reminder sending, random assignment, automatic attribution and statistical calculation have not been performed here. Any external contact requires its specific approval.
How to compare
- Before starting, define the eligible population, assignment unit, metric, window, exclusions and decision date. If assignment is by customer, keep that customer's other checkouts out of the other group.
- Random assignment can help compare groups, but requires appropriate design, sample size and control of other contacts or offers. Retrospectively labeling buyers and non-buyers is not enough.
- Maintain comparable order observation in both groups. Document cancellations, refunds and costs if they are part of the objective; do not invent net revenue from order values alone.
- Do not stop at the first favorable result. Review incidents or harm and apply the agreed evaluation plan. If evidence is insufficient, acknowledge that the result is inconclusive.
Waitlist; no promise of recovery, savings or integration. Spanish destination.
Frequently asked questions
Is a recovered order in a report an additional sale?
Not necessarily. It may satisfy the report's attribution rule without establishing what the person would have done without the message. Preserve that metric's name and scope.
Can I divide purchases by messages sent?
That describes a relationship only with a clear unit and window; it is not customer conversion or incremental effect. For a group comparison, fix the denominator of eligible assigned records and the observation before sending.
Does this example prove reminders do not work?
No. It is fictional arithmetic to avoid inflating results. Neither a positive effect nor a null effect has been established for your store.
Prepare cart recovery (Spanish)
Sources
- Shopify — abandoned checkouts
- Think² reference (Spanish)
- Think² reference (Spanish)
- Think² reference (Spanish)
Spanish destination.