First, match tracked conversions to orders by reference and classify every gap
Begin where the attribution and deduplication guide begins its reconciliation: the order system, the only complete record. Export the affiliate platform's conversions for one complete period, with order reference, partner, order value, status and the date, and match them to orders by reference. The matches are the population you can explain; the conversions that match nothing are the first class of gap, and usually duplicates, test orders or postbacks sent without a reference. Then take each matched conversion and ask what analytics did with the same order. The guide's classes are the ones this page uses: credited to affiliates by both systems, credited by the platform while analytics names another channel, credited by the platform with no session in analytics, rejected at validation, or returned afterwards. A sixth class belongs to partners rather than channels: orders where analytics shows the session arriving from one partner's site while the platform paid another, because a click at the checkout overwrote the first.
Before comparing anything, write down the rule each system applied. The attribution models framework is the vocabulary: the platform runs a last-affiliate-click rule inside its own window over the touchpoints it alone can see; analytics runs the model the property is set to, over a lookback the property sets, across every channel it recorded. The attribution guide notes that analytics tools document their own attribution settings and lookback windows, and that Google Analytics Help describes how its reports assign credit and how the settings can be changed, so know the settings in your own property before you compare. The guide to platform versus analytics discrepancies adds the discipline that keeps the comparison honest: compare like with like, reconcile complete periods only after both systems have finished processing, set both sides to the nearest equivalent view, and treat the residual as something to document rather than to hide.
Three neighbouring pages are linked rather than repeated. The diagnosis of channel conversions adding up to more than sales covers every platform's claims together exceeding the orders taken; the tutorial on tracing missing affiliate conversions runs in the opposite direction, from a partner's claim that sales are missing to a tracking fix; and the attribution guide owns the reconciliation table itself and the question of which figure to present to finance. This page is for the single gap between the affiliate platform and analytics, read one tracked conversion at a time.
Diagnosis table
Diagnosis
| Symptom | Check | Likely cause | Fix |
|---|---|---|---|
| Both systems record the order, but analytics credits it to a channel the platform cannot see, and the two totals drift apart whenever a setting changes | The platform's window and rule against the property's model and lookback, and the date of the last change to either | The platform credits every affiliate click within its window while analytics applies its own model across all channels | Record both rules, compare under the nearest equivalent view, and document the expected gap as a baseline in the measurement plan |
| The platform's total is quoted before validation while analytics and the order system show orders that were returned, cancelled or rejected | The status of each tracked conversion: pending, approved, rejected, returned | Tracked conversions compared with validated ones, before returns, cancellations and rejections | Compare approved, deduplicated conversions with analytics, and report tracked and paid as separate figures |
| A share of tracked conversions matches no order at all | The reference field in the postback and the platform's duplicate handling | No order reference in the postback, so lines cannot be matched to the order system | Carry the order reference in every postback, let the platform deduplicate pixel and postback, and reject the unmatched |
| Orders match, the platform credits a partner, and analytics has no session for the customer | The property's consent rate and the three consent states on a real page | Visits where consent was declined or the analytics tag was blocked, invisible to analytics but tracked server-side | Size the class against the consent rate, confirm the tags behave as designed in each state, and keep the affiliate setup consistent with the consent rules |
| Orders match, analytics credits paid search or email, and the affiliate click came earlier in the path | The session source analytics recorded and the timing of the later click against the affiliate click | Paid search and email clicks after the affiliate click, so analytics credits the later channel | A written cross-channel deduplication rule in the terms, with last click on both sides as the comparison view |
| Analytics shows the session arriving from a content partner's site while the platform paid a voucher site or an extension | The partner credited by the platform against the referring site in analytics, and the time between the two clicks | A partner's click overwritten by an extension or voucher partner at checkout | Rules in the terms on extensions and codes, deduplication that protects the earlier partner, and a partner-level comparison each month |
Fix, verify and prevent each cause
Different windows and rules on each side
The attribution and deduplication guide explains why the platform's total is always the larger. The platform credits the most recent affiliate click inside its cookie window and is blind to everything else on the path, whether a paid search ad, an email or a direct visit, before or after that click, so by its own rule any order an affiliate touched counts as an affiliate order. Analytics credits one or several channels under its own model, and the guide's table of rules shows that a cross-channel last click, a first click and a data-driven split give different answers to the same path. The comparison of attribution models adds a detail that easily trips a reconciliation: the platform's last click and analytics' last click are not the same rule, because the analytics version ignores direct arrivals and the platform's version ignores every other channel.
Fix: record the window and the rule on each side and the dates they last changed, then set the comparison to the nearest equivalent view, which the models comparison names as last click on both sides because no other rule comes out of every tool in the same shape. Treat the orders that remain differently credited as the expected consequence of two different questions, and document the size of that class as a baseline, which the discrepancies guide places in the measurement plan's source-of-truth table.
Verify: on the next complete month, the share of matched orders in this class should sit inside the documented range; a jump without a change in partner mix points at a setting that moved.
Prevent: keep one analytics model constant for any series reported over time, date every change to a window or model on either side, and never let a change appear in the numbers before it has been explained to the people reading them.
Tracked compared with validated
The tutorial on closing the month on affiliate commissions lists tracked, pending, approved, rejected, invoiced and paid as different figures and calls a tracked conversion a claim rather than a cost. A reconciliation that puts the platform's tracked total beside analytics compares a number that still contains returns, cancellations, duplicates, test orders and conversions the terms exclude with a number that never contained most of them. The attribution guide draws the same line for reporting: base the figures on validated conversions, because tracked ones flatter the programme's return on spend and cost per action.
Fix: validate every pending conversion against the order system by reference before the platform's approval date, hold conversions whose returns window is still open as pending, and compare approved, deduplicated conversions with analytics rather than the tracked total. Keep chargebacksA sale reversed by the customer's card issuer, which can arrive after other checks have closed and should not stay in payment or attribution figures. that arrive after approval in a clawback line of their own. Report tracked and paid as two figures with what each answers.
Verify: the month's approved total should reconcile to the order system's qualifying orders, and the gap to analytics should shrink by the rejected and returned classes exactly, which the tutorial's reconciliation step records.
Prevent: make the close a fixed monthly routine with a named approver, keep the validation periodHow long the business has to review an affiliate conversion before approving its commission, long enough to cover returns and cancellations. in the terms long enough to cover the returns window, and never quote the tracked figure to finance without the paid figure beside it.
No order reference in the postback
The guide to how affiliate tracking works describes a postbackA server-to-server message reporting a conversion, install or in-app event back to the ad network or affiliate platform that should receive credit. as the advertiser's server telling the platform that an order was placed, carrying the click identifier stored earlier and the order details, and it names the order reference as what makes validation workable, because platform records can then be matched to the order system automatically rather than by hand. The attribution guide is blunter: without an order reference in postbacks, reconciliation becomes guesswork. A programme whose postbacks carry value and status but no reference cannot tell a duplicate from a second order, cannot tell which tracked conversion was returned, and cannot build the classes above at all; the glossary entry describes the mechanism. The tracking guide also notes that many advertisers run a postback and a browser pixel together with the platform removing duplicates, and where that removal fails, the same order is tracked twice.
Fix: add the order reference to every postback, send status updates for returns and cancellations through the same route, let the platform deduplicate pixel and postback on the reference, and reject tracked conversions that match no order. Send only what the platform needs: the click identifier, the order reference, the value and the status, and never personal data.
Verify: export a month and match by reference; every approved conversion should match exactly one order, and the unmatched class should be empty apart from deliberate test orders.
Prevent: treat the postback as an integration the business owns and monitors, test it after every checkout change, and alert on a sudden rise in unmatched conversions, which is how a reference field that stopped populating shows itself first.
Orders analytics never saw
The discrepancies guide explains the class that no setting can close: a visitor who declines analytics storage is never observed by analytics, so a market where few people consent shows an analytics shortfall that says nothing about how campaigns performed. Ad blockers and privacy tools remove more sessions in the same way. The affiliate platform, where the setup stores the click identifier first-party and records the sale by postback, can still match the order to the click, so the platform credits a partner for an order that has no analytics session. The tracking guide adds the constraint that keeps this class honest: affiliate tracking is usually treated as non-essential, so where consent is required the identifier should not be stored or shared until the customer has agreed, and the consent banner and the tracking setup must agree.
Fix: size the class against the consent rate the consent platform reports, as the diagnosis of a data drop after a cookie banner does, treating declined visitors as the expected loss and anything beyond it as an implementation error to test on a real page in each consent state. Confirm the affiliate setup honours the same consent rules rather than firing regardless, and tell partners in the terms that some conversions will never be attributable.
Verify: the share of matched orders with no analytics session should move with the consent rate by market and nothing else; a share larger than the declined share means a tag is blocked or mistimed rather than declined.
Prevent: keep the consent rate beside the reconciliation table every month, retest the three consent states after every banner or tag change, and record the result, so a widening gap in this class is read as a consent change rather than as a tracking fault or a partner's doing.
A later paid search or email click
The comparison of attribution models describes last click as the rule most ad platforms and finance teams understand, and notes what it rewards: brand search, retargeting and email sit at the end of the journey, so under a cross-channel last click an affiliate click followed by a brand search click or an abandoned-basket email loses the order to the later channel. The affiliate platform cannot see that later click, so it credits the partner; analytics can, so it credits search or email; and both are applying their rule correctly. The attribution guide turns the class into a decision: deduplication is the choice of which channels override an affiliate click, within what period before purchase, and it belongs in the programme terms before it is applied, with UTMURL tags identifying traffic source. parameters on paid and email links so the site can tell the channel of each visit.
Fix: write the cross-channel rule into the terms, naming the overriding channels and the period, and apply it at validation or at the moment the postback is fired; for the comparison itself, set both sides to last click so that the orders in this class are the ones the rule was written for. Keep the payment rule and the analysis separate, because the guide warns that last click is a contract, not a measure of value.
Verify: the orders analytics credits to search or email after an affiliate click should be exactly the orders the deduplication rule removes or keeps, as the rule says; a partner should be able to check its own statement against the rule.
Prevent: keep the paid search team from adding affiliate parameters to their own links and partners from passing search traffic through affiliate links, as the guide asks, and explain any change to the rule to partners before it appears in their numbers.
A checkout click overwrote the partner
The recruitment, compliance and fraud guide describes the mechanism: some extensions detect a shopper on a checkout page and offer a code or cashback, inserting their own tracking link at the last moment, and when that overwrites the click of the partner who sent the customer, the platform pays the wrong partner. The same guide covers the voucher version, where a code issued to one partner leaks and a shopper searching for it at the checkout gives a voucher site the last click, which the diagnosis of coupon sites taking affiliate commission treats as a problem of its own. In the reconciliation, the order matches and the counts agree, but analytics shows the session arriving from a review site while the platform credits an affiliate browser extensionA partner whose software sits in the shopper's browser and pops up at checkout with codes or rewards, often taking the last click on an order already decided. or a voucher site, and the review site's analytics-credited orders exceed its platform orders month after month.
Fix: write rules in the terms that restrict extensions to activations the customer chooses and prohibit automatic link insertion, tie every code to the partner it was issued to and withhold commission when a code appears through someone else, and set the between-affiliates deduplication rule so that protected content partners keep their credit when a checkout partner clicks last, as the attribution guide's rule table allows.
Verify: compare partner-level figures each month, the platform's credited partner against the referring site in analytics for the same orders; the class should fall to the cases the rule deliberately allows.
Prevent: keep the list of protected partners and permitted extension behaviour visible in the terms, review the overrides, and treat a rising share of commission to checkout partners with a flat new-customer share as the early sign, which the diagnosis of affiliate commission costs rising reads as a cost problem before it becomes one.
Prevention: an order reference in every postback and a monthly reconciliation table from tracked to paid
The six classes above only became an argument because nobody could build them. The habit that prevents the next argument is a reconciliation produced every month from the same export the close already uses, running from tracked conversions to paid conversions line by line, with each step given its reason. The attribution guide's version of the table ends most of the disputes between finance, the affiliate team and the paid search team because every line has a cause, and the measurement hub places that table among the programme's other measurement tasks: who deserves credit, what each partner delivers and what the programme adds.
The reference is the foundation. Every postback carries the order reference, the click identifier, the value and the status, and nothing else; status updates for returns and cancellations travel the same route; and the platform deduplicates pixel and postback on the reference. With that in place, the match to the order system is automatic and the unmatched class is the first line of the table rather than a mystery. The postback is an integration the business owns, so it is tested after every checkout change and watched for sudden drops in recorded conversions per partner, as the tracking guide asks.
The table is the routine. For each complete period, after both systems' processing has finished: tracked conversions from the platform; less the unmatched; less rejected and returned, from validation; the remaining matched orders split by what analytics recorded, the affiliate channel, another channel after a later click, or no session; the deduplication rule applied to the later-click class as the terms state; the partner-level check for checkout clicks that overwrote an earlier partner; and the paid figure at the bottom, reconciled to the order system and the invoice. Beside it sit the settings that explain the expected classes: the platform's window and rule, the property's model and lookback with the date they last changed, the consent rate by market, and the documented baseline range for each class, so that a month outside its range is read as a question about one class rather than as a collapse of the channel.
Three figures go to the people who need them, with what each answers. Tracked conversions are what partners touched. Paid conversions are what the business paid commission on after validation and deduplication, and the figure finance books. Incremental conversions, where the programme has run tests, are what it added beyond what would have happened anyway. Showing all three, and the trend in each, is the attribution guide's answer to the question of which figure to report, and it stops the analytics number being quoted as a count of affiliate sales when it was only ever a view across channels under one model. Explain any change to a rule, a window or a consent banner before it appears in the numbers, and keep the deduplication rule, the extension rule and the list of protected partners in the terms where partners can read them. When the question widens from this one pair to every platform's claims against the orders taken, the diagnosis of channel conversions adding up to more than sales takes over, and the affiliate troubleshooting hub lists the neighbouring symptoms, several of which are this table read from a different line.
A camera accessories shop explains a quarter's gap line by line (illustrative)
Prenduvax Camera Supplies
Prenduvax Camera Supplies, an invented online camera accessories shop, is asked by its finance director why the affiliate network reported 3,140 conversions for the quarter while Google Analytics credits 1,790 orders to the affiliate channel. The programme manager exports the quarter's conversions with order references and matches them to the order system: 3,095 match an order, and 45 do not: 42 duplicates from a week in which the network's pixel and the shop's postback both fired, plus three test orders. Those 45 are rejected.
Of the 3,095 matched orders, analytics credits 1,790 to affiliates, 620 to paid search or email after a click that came later than the affiliate click, 375 to direct or organic, and 310 have no analytics session at all; the consent platform reports that 26% of visitors declined analytics storage in the quarter, which fits. At validation, 265 of the matched orders were returned or cancelled inside the 45-day window. The partner-level comparison finds 150 orders where analytics shows the session arriving from one of two camera review sites while the network paid a voucher extension whose click arrived less than two minutes before checkout.
The shop writes its rules into the terms: paid search clicks within 12 hours and email clicks within 6 hours of purchase override an affiliate click, and an extension click within 30 minutes of checkout does not overwrite an earlier content partner's click. Applying the first rule deduplicates 360 of the 620 later-click orders; the second reassigns the 150 to the review sites. The reconciliation table runs from 3,140 tracked, less 45 unmatched, less 265 returned, less 360 deduplicated, to 2,470 paid conversions, with the 310 consent-declined orders and the 375 direct and organic orders shown as explained classes rather than losses. Finance books commission on 2,470 orders, accepts the paid figure beside the tracked one, and the quarterly table now carries both with the consent rate and the settings that explain the gap.
Common mistakes
- Picking a winner. The platform and analytics answer different questions; the order system is the only complete count, and the table explains the rest.
- Comparing the tracked total. It still holds returns, duplicates and test orders that analytics never counted.
- Postbacks without an order reference. Every class above depends on matching by reference; without it, reconciliation is guesswork.
- Treating consent loss as a tracking fault. The declined share is expected; size it before touching the tags.
- Comparing the platform's last click with analytics' model. Set both sides to last click first, then explain the residual.
- Deduplicating in secret. Rules applied without being in the terms turn every statement into a dispute.
- Ignoring the partner level. Counts can agree while the wrong partner is paid for the order.
- Reconciling too soon. Compare complete periods after both systems have finished processing, or the gap closes by itself and the fix was for nothing.
