How it worksReadiness scoreWhy PingPlusLearnBlog
Get started freeLog in
Skip to guide

Practical guide · Part 03

Validate AI Advertising Conversions: Duplicates and Refunds

A campaign report can record the right event while the business receives fewer useful results. Reconcile the event with the record that owns the outcome, and keep billing, qualification, attribution, and value separate.

PingPlus Editorial TeamReviewed

Reconcile a worked booking ledger

Use one row per relevant event or outcome identifier and retain the reason for every adjustment. Do not silently delete inconvenient records.

Worked ledger · synthetic $240 cohort

Ten events do not mean ten customer outcomes
E08E09 duplicate
B08One booking, counted once
  1. 10Recorded eventsKeep every raw row
  2. 9Unique bookingsE08 + E09 share B08
  3. 6Accepted bookingsExclude cancelled, unsupported, and test records

5 completed$240 ÷ 5 = $48 per completed job so far

1 pendingAccepted; awaiting completion

Deduplicate 10 events into 9 bookings, validate 6 accepted bookings, then separate 5 completed jobs from 1 pending job; $240 divided by 5 gives $48 per completed job so far.

One accepted booking is still pending. Business-quality exclusions do not automatically change billing; apply the actual campaign terms.

Start by comparing a normal outcome with a repeated one. These synthetic records belong to the ten-event, $240 cohort below.

Normal and duplicate records
RecordHow to count itEvidence
E01 → B01; accepted and completedOne raw event, one accepted booking, one completed jobE01 and the status history of B01.
E08 → B08; accepted and completedOne raw event and one accepted, completed bookingE08 and the status history of B08.
E09 → B08; repeated outcomeKeep the raw event; add no second accepted booking or completed jobE09 points to the same B08 business record as E08.

The full synthetic cohort contains ten recorded booking-request events and $240 in campaign spend. It is a teaching dataset, not an observed customer result. The illustrative business definition counts unique, eligible, accepted bookings; completed jobs are a later stage.

Ten synthetic booking-request events
EventBusiness recordReview result
E01B01Accepted and later completed.
E02B02Accepted and later completed.
E03B03Accepted and later completed.
E04B04Accepted and later completed.
E05B05Accepted; appointment still pending.
E06B06Cancelled before acceptance.
E07B07Unsupported postcode; rejected.
E08B08Accepted and later completed.
E09B08Repeated event for the same booking; duplicate.
E10B09Test request; excluded from customer outcomes.

The cohort contains six accepted bookings: B01–B05 and B08. Five are completed; B05 is still pending. The ten event rows are not ten unique bookings, and the five completed jobs should not be confused with the six accepted bookings.

Use the event ID to identify repeated delivery of the same event, then the business record ID to reconcile the customer result. E08 and E09 have different event IDs but point to the same B08 booking, so the stated outcome is counted once. Two genuine bookings by one customer can have different booking IDs and both count. Do not deduplicate solely by email, customer name, or a similar timestamp.

Use the actual campaign terms to determine how E06, E07, E09, and E10 affect billing. This example only shows how they affect the stated business-quality counts.

Define the event and its business meaning

Write down what was recorded, what makes it useful, where it is confirmed, and which billing rules apply. Those definitions may differ.

A submitted booking request may be the selected action, while the business ultimately values an eligible accepted booking or a completed appointment. A confirmed sale can later be refunded. Start with those states before comparing platform and business totals.

Definitions to agree before reviewing results
DefinitionQuestion to answerExample
Recorded eventWhat exactly triggered the reported action?A booking request reached the confirmation page.
QualificationWhich conditions make the result useful?Supported service area, real customer, unique request, accepted time.
Source of confirmationWhich business record owns its status?The scheduling system and its booking ID.
AttributionWhich supported rule assigns campaign credit?The documented method and review window for this test.
Billing treatmentWhat does the campaign agreement charge or adjust?The agreed event and reconciliation terms; do not infer them from a generic guide.

These are measurement recommendations, not universal PingPlus rejection, refund, or payment policies. Confirm the campaign's actual terms before activation. An event excluded from a quality metric is not automatically a billing credit.

Calculate the cost at each useful stage

Divide the same clearly defined spend by the outcome count for the stage you are evaluating. Name the stage in the metric.

Worked cost calculations
MetricCalculationResult
Cost per recorded request event$240 ÷ 10$24
Cost per accepted eligible booking$240 ÷ 6$40
Cost per completed job so far$240 ÷ 5$48
Pending accepted bookings6 accepted − 5 completed1; outcome not mature yet

A $24 headline does not describe the cost of a completed job. Equally, the $48 figure is provisional while an accepted appointment remains pending. Review the same acquisition cohort again after its normal completion window.

For a fictional sales example, four $100 orders produce $400 gross order revenue. One is fully refunded, leaving $300 before product, fulfillment, payment, and other costs. If campaign spend is $120, gross-order ROAS is 3.33 and revenue-after-refund ROAS is 2.50. Neither ratio alone establishes profit. State the revenue basis and review window.

Investigate differences without forcing the totals to match

Check event definitions, time windows, identifiers, and later statuses before treating a difference as a reporting error.

  1. Align the period. Compare the same timezone, event date, and cohort basis. A creation-date report can differ from an appointment-date report.
  2. Join identifiers. Find whether repeated events point to one order or whether a reschedule created a new record. Preserve the relationship rather than counting both as new customers.
  3. Review changed status. Confirm when a booking was accepted, cancelled, completed, or refunded. An earlier export can legitimately differ from today's status.
  4. Check attribution separately. A real order without supported campaign evidence can remain a business sale without receiving campaign credit.
  5. Resolve the billing question through the agreed process. Keep the record, reason, evidence, and resolution. Do not manually rewrite the ledger to make it equal the invoice.

Use non-sensitive internal identifiers and limit access to the people who reconcile outcomes. A shared report should not expose customer contact details merely to show that two events were duplicates.

Test duplicates, missing records, and later adjustments

Give each test a business state to inspect and a reporting result to expect. Verify the actual record before calling an event a qualified conversion.

Run a small set of marked test journeys before relying on the report. The cases below are separate test scenarios, not additional rows in the ten-event ledger above.

Conversion validation test cases
ScenarioExpected business and reporting resultEvidence to retain
The confirmation page is refreshed or event delivery retriesPreserve the same event identity for a retry; the customer result is counted onceEvent IDs, business record ID, and the collection path that caused the repeat.
Two event IDs point to the same bookingRetain both raw records, flag the repeated outcome, and count the booking onceBoth event IDs and the shared booking ID; the agreed qualification rule.
One customer makes two legitimate bookingsRetain two distinct results when both satisfy the definitionDifferent booking IDs and evidence that both bookings are genuine.
An event exists but no matching business record can be foundMark the result unresolved; do not assume acceptance from an analytics eventEvent time, identifiers, investigation owner, and resolution status.
A booking is accepted and later cancelledRetain its status history; update the metric according to the stated outcome definition and review dateAcceptance and cancellation times, current status, and applied reporting rule.
A separate $100 test order receives a $40 partial refundRetain the order and record $60 after refund before other costs; identify the value basis in reportingOrder ID, refund ID, currency, amount, and adjustment time.

Separate a business correction from a platform adjustment

Updating a CRM or internal ledger does not automatically update an ad platform's conversion report. Check whether the actual channel supports an adjustment and how it identifies the original conversion. For example, Google Ads distinguishes a restatement from a retraction: a restatement changes value, while a retraction removes a conversion. These are Google Ads mechanisms, not a claim that every channel or PingPlus campaign uses the same interface or rules.

Record the requested adjustment, response, and later report result. A queued or failed request stays unresolved until verified. Keep the selected campaign's billing treatment separate from conversion-report corrections.

Save each review with its cohort, event definition, time zone, reporting cutoff, spend basis, and pending outcomes. Revisit the same cohort after normal completion or refund windows instead of comparing an immature week with a settled one.

Choose a campaign action from the findings

Fix the earliest material source of wasted effort or spend, then review comparable results before expanding.

In the booking example, investigate the unsupported postcode, duplicate event, and test request. The first is a fit or validation issue; the second is a measurement issue; the third needs a reliable exclusion process. Those problems require different fixes.

Continue when the accepted outcomes meet the business's cost and capacity goals. Adjust the offer or path when the problem is identifiable. Pause when the team cannot trust the records or serve the demand. Increase one variable at a time after the basic economics and validation hold.

A small cohort can expose a broken process. It cannot establish that one channel is universally cheaper or that every accepted booking was incremental. Use the campaign measurement guide to plan the next review, and performance models to compare the selected billing events.

Keep a complete reconciliation record

Retain the identifiers, status history, amounts, and evidence needed for someone else to reproduce the decision.

Reconciliation material checklist
MaterialWhat to retain
Event and business identityCampaign and event IDs; booking, lead, or order ID; links between repeated or replacement records.
Definitions and cohortSelected action, qualification rule, attribution rule and window, time zone, reporting cutoff, and spend basis.
Status and timesCreation, acceptance, completion, cancellation, and adjustment times as applicable; current state and previous states.
AmountsOriginal amount and currency, full or partial refund amount, adjusted value, and any pending value. Label the revenue basis.
Reason and evidenceDuplicate relationship, unsupported area, test marker, cancellation or refund reason, and reference to the supporting business record.
ResolutionOwner, decision date, outstanding check, requested platform adjustment, response, and verified report result.

For E09, keep the link to B08 and the duplicate reason; the raw event remains visible while the accepted-booking count stays unchanged. For a missing business record, keep the item unresolved with an owner. For the separate $100 order with a $40 partial refund, retain both amounts and report $60 after refund before other costs. Apply the agreed billing process separately from those reporting decisions.

Close a review when another reviewer can follow the identifiers to the business evidence, reproduce the counts and values at the stated cutoff, and see which platform adjustments are verified or still pending. Share internal references with appropriate access rather than copying customer contact details into the report.

FAQ

Does a duplicate conversion automatically qualify for a refund?

Not automatically. Your campaign agreement and validation process determine how duplicates affect billing or refunds. Keep the evidence, apply the agreed outcome definition, and use that process to reconcile the record.

Should cancelled bookings disappear from reporting?

Keep cancelled bookings in the record with their current status and dates. You can leave them out of a clearly defined qualified-outcome metric while keeping them visible in raw events and cancellation reporting. That makes it easier to explain how you reached your totals.

Which CPA should I show the business?

Show the cost at each stage your team uses: recorded actions, accepted qualified outcomes, and completed customer results where available. Include the spend, counts, outcome definition, and how much time results have had to mature, so the numbers are easy to interpret.

What if campaign spend is positive but there are no qualified outcomes?

Show the spend and zero qualified outcomes. Cost per qualified outcome can't be calculated as a finite amount here, so don't report zero CPA. Check customer fit, action completion, measurement, and follow-up before increasing spend.

Should I deduplicate by customer or by conversion?

Deduplicate by the outcome you're counting. Two records for one booking should count once; two genuine bookings from the same customer can count twice. Keep both event and business record identifiers so you can check the difference. A customer identifier alone doesn't establish whether the outcome is a duplicate.

Continue with your next task

Sources and further reference

The worked examples are teaching materials. Apply the method to your actual offer, records, and campaign terms.