Different questions, different numbers
Every reporting system is built to answer the question its owner cares about. An advertising platform is built to show an advertiser what their spend produced, so it counts every outcome it can connect to an exposure to its ads — clicks, and in most platforms views — within a window it sets, using identifiers it holds, including matches it has modelled where consent or browser restrictions prevented direct observation. An analytics tool is built to describe what happened on the site, so it counts events the site sent it, attributes each across every channel it can see under the property's AttributionAssigning conversion credit to touchpoints. model, and knows nothing about exposures that did not produce a visit.
Neither view is wrong. The platform's number is the answer to "how much did my ads influence?", as the platform defines influence. The analytics number is the answer to "where did the visitors who converted on the site come from?", as analytics defines coming from. The business's own order system answers a third question — "how many orders were there?" — and is usually the only one of the three that counts the whole population.
The practical consequence is that reconciliation is not a search for the true number. It is the work of listing the differences between the questions, estimating what each contributes to the gap, and deciding in the measurement plan which number informs which decision. The platform attribution versus analytics comparison covers the attribution half of that decision in depth; this guide covers every cause and the method.
The causes
The table is the diagnostic reference. Each cause has a direction it pushes the platform figure relative to analytics, and a place to look.
| Cause | Effect | Where to check |
|---|---|---|
| Clicks counted by the platform, sessions by analytics | Platform clicks exceed analytics sessions | Click-to-session ratio by campaign; landing page load and tag timing |
| View-through conversions | Platform conversions exceed analytics conversions, most on display, video and social | Platform conversion settings: whether views count and in which window |
| Attribution model and lookback window | Platform credits its own ad fully within its window; analytics shares credit across channels and its own window | Platform attribution setting; analytics attribution settings |
| Conversion date assignment | Platform may report on the date of the click; analytics on the date of the event | Date-of-conversion versus date-of-interaction reporting options |
| Consent and tag blocking | Analytics loses declined visitors; platform may model them | Consent rate by market; consent mode form; modelled-conversion indicators |
| Modelled and matched conversions | Platform includes modelled, enhanced or server-matched conversions analytics never saw | Platform documentation for the conversion source; enhanced conversions and server events status |
| Tagging and UTMURL tags identifying traffic source. errors | Analytics assigns paid traffic to organic, referral or direct | Channel report for the campaign; redirect handling; auto-tagging status |
| Cross-device and cross-browser | Platform matches a signed-in user across devices; analytics sees separate users unless an identifier is passed | Identity settings in analytics; platform reporting on cross-device |
| Filters, segments and thresholds | Analytics excludes internal traffic or applies privacy thresholds; platform does not | Filters active on the property; "(other)" rows and threshold notices in reports |
| Time zone | The two systems' days start at different hours | Property time zone against platform account time zone |
| Processing latency | The platform keeps attributing conversions for days; analytics finalises on its own schedule | Reporting date against today; documented processing windows for each |
| Duplicate or misconfigured conversion actions | Platform counts one action under two definitions, or counts every instance rather than one per click | Platform conversion action list and counting setting |
| Bots and invalid traffic | Platform filters invalid clicks before billing; analytics may still record some sessions | Invalid-click reporting; bot filtering settings |
Work the table from the top for a click-and-session gap and from the view-throughA conversion credited to an ad that was shown but not clicked, within a set view window. row for a conversion gap; the causes nearer the bottom usually explain the residual once the large ones are accounted for.
When platform figures are copied into a warehouse, the latency and attribution rows become loading problems too, and the guide to loading ad platform data into a warehouse covers the trailing reload window and the settings history that keep loaded numbers comparable with the interfaces.
Counting units: clicks, sessions, users, conversions
Most confusion starts with units. A click is a platform event: the ad was tapped. A session is an analytics construct: a period of activity on the site that begins when the tag first runs and ends after a documented period of inactivity or at a day boundary, depending on the tool's rules. A user in analytics is an identifier — a browser-level cookie, or a signed-in id where the site passes one — and a user in a platform is an account the platform recognises across devices. A conversion in the platform is any outcome it could connect to an ad; a conversion in analytics is an event marked as key, counted once per event or once per session according to its setting.
Clicks exceed sessions for mechanical reasons: the page did not finish loading, the visitor tapped twice, the tag was blocked or waited for consent, the click was invalid and later removed from billing but not from the click count, or the landing page redirected and the parameters were lost. The ratio should be stable per campaign and device; a sudden change in it is a landing page or tag problem, and the data quality guide has the weekly check that catches it.
Conversions differ for a different reason. The platform counts outcomes connected to exposures, so a person who saw an ad, did not click, and bought a day later through a search is a platform conversion and, in analytics, an organic search conversion. The platform counts once per click or every instance depending on its setting; analytics may count once per session. Users differ because the platform can see one person on two devices where the site cannot, and because analytics counts a new identifier after cookies are cleared.
Compare like with like. Clicks against sessions, platform conversions against analytics conversions filtered to the same channel, and both against the order system. Comparing platform conversions with all analytics conversions, or clicks with users, produces a gap that has no cause because the two numbers were never measuring the same thing.
Attribution and windows
Even with identical units, the two systems give credit differently. A platform typically credits a conversion to its own ad if the person clicked, or in many platforms merely saw, that ad within a lookback window the advertiser sets from the documented options, and it does not know about the other channels the person touched. Analytics sees every channel that produced a visit and distributes credit by the property's model — a data-driven model by default in GA4Google's event-based analytics platform., with rule-based alternatives documented — across a lookback window set in the property.
Three consequences. Platforms' conversion totals summed across every platform exceed any single count of orders, because each platform claims the conversions it touched and several touched the same ones. Analytics attributes a smaller share to any one paid channel than the channel's own platform does, because credit is shared. And changes to either window or model move the numbers without any change in marketing, so the measurement plan records both settings and the date of every change.
The attribution-models framework and the platform attribution versus analytics comparison cover the choice of model. For reconciliation, the task is narrower: note the window and model on each side, set the comparison to the nearest equivalent each tool allows (a last-click, click-only view on both sides is the common starting point), and treat the residual as the influence of views and shared credit. Where the question is whether the channel caused the conversions at all, neither system answers it; an IncrementalityConversions caused by marketing. test does.
Consent, modelling and privacy thresholds
Consent affects the two systems unequally. Where analytics storage is declined, analytics does not observe the visitor at all, or observes only cookieless signals under the advanced form of consent mode and models the aggregate where the documented conditions are met. Where advertising storage is declined, the platform cannot match the conversion to the click by cookie, but may still match it through a first-party identifier sent with enhanced conversions or a server-side event, or may model it. The result is that a market with a low consent rate shows a large analytics shortfall and a smaller platform shortfall, and the gap between the systems widens for reasons that have nothing to do with campaign performance.
Modelling adds a second layer. Platform reports include modelled conversions that are estimates rather than observations; analytics includes modelled users and events where behavioural modelling is active. Both are documented, both are labelled in the interfaces, and neither is reconcilable to an order system because there is no event behind the estimate. The reconciliation method below separates observed from modelled before comparing.
Privacy thresholds are a third effect specific to analytics: Google documents that some reports withhold rows that could identify individuals when certain features are active, which can remove small segments from a report entirely. A campaign report that seems to miss a few conversions may be withholding them rather than lacking them; the interface indicates when thresholds have been applied. The privacy and consent guide covers the design response.
Timing, time zones and latency
The two systems keep different clocks. A platform account and an analytics property each have a time zone setting, and when they differ, every daily figure is offset by the difference; a conversion late in the evening lands on one date in one system and the next in the other. Set them to the same zone where the business allows, and record it in the plan where it does not.
They also assign dates differently. Many platforms report a conversion on the date of the click that earned it, so a report for last week keeps changing as conversions from last week's clicks arrive; analytics reports the event on the date it happened. A week-old platform report and a week-old analytics report therefore describe different things, and a comparison run too soon will show a gap that closes by itself. Google documents the processing windows for each product; the plan should say how many days to wait before a period is treated as final, by linking to the documentation rather than fixing a number that may change.
Latency also runs the other way: analytics standard reports are updated on a documented schedule, and an exploration run on a partially processed day will differ from the same report tomorrow. The rule is simple: reconcile complete periods only, after both systems' documented processing has finished, and never compare a figure read this morning with one read last Friday.
A reconciliation method
Reconciliation is a repeatable procedure that ends with a documented, stable difference and a decision about which number to use. Run it once to establish the baseline and then monthly to confirm the baseline holds.
- Choose the pair and the period. One platform, one analytics property, one complete month after both systems' processing has finished.
- Align the clocks. Confirm time zones match or record the offset. Confirm the date basis — date of conversion on both sides where the platform allows it.
- Align the units. Clicks against sessions from that platform's channel; platform conversions against analytics key events filtered to that channel; both against the order system for the same period.
- Align the rules. Set both sides to the nearest equivalent attribution view: click-only, last click, matching windows where possible. Record what could not be matched.
- Separate observed from modelled. Where either side labels modelled figures, compare observed to observed first and note the modelled share separately.
- Attribute the gap. Work down the causes table, estimating each cause's contribution: view-through share from the platform's own breakdown, consent loss from the consent rate, tagging loss from the channel report, duplicates from the conversion action list. Stop when the residual is small and stable.
- Fix what is a fault. Tagging errors, duplicate conversion actions, time zone mismatches and broken tags are data quality findings, not discrepancies; raise them and re-run after the fix.
- Document the baseline. The expected ratio and its usual range for this pair, the causes and their shares, the settings on both sides, and the date. Put it in the measurement plan's source-of-truth table, which the measurement planning hub places as the last artefact in the sequence.
- Decide the use. Which number informs bidding, which informs reporting to the business, which informs budget allocation, and where an incrementality test is needed because neither system can answer.
Reconciling a paid social campaign for one month
A subscription box company compared its paid social platform with GA4 for May. The platform reported 12,400 clicks, 1,180 conversions; GA4 reported 9,900 sessions from that channel and 610 key events; the billing system recorded 2,050 new subscriptions from all channels. Step by step: the click-to-session ratio of 0.80 matched the previous 3 months, so no tagging fault. The platform's own breakdown showed 410 of the 1,180 conversions were view-through, leaving 770 click-through. The consent rate for the main market was 71%, so GA4's observed 610 implied roughly 860 if scaled, within range of the 770 once the platform's 7-day window against GA4's last-click model was allowed for. A duplicate conversion action was found counting both a page view and a server event for the same subscription, inflating the platform figure by about 90; it was fixed. The documented baseline became: platform click-through conversions exceed GA4 by between 15% and 30%, driven by consent and cross-device matching; view-through is reported separately; billing is the figure for the business; and a geo-holdout test was scheduled because the channel's incremental contribution was the open question.
When comparing either figure with published numbers, the guide to using industry benchmarks explains why definitions and attribution windows make most comparisons unsafe.
When the question is why the conversions claimed by all channels together exceed the orders taken, channel conversions adding up to more than sales diagnoses shared credit, view-through, windows, dating, duplicates and refunds.
When an affiliate platform reports far more sales than analytics credits to the channel, the diagnosis of affiliate sales not matching analytics applies the same classification to one pair, from tracked conversions matched by order reference to the paid figure.
When Campaign Manager credits LinkedIn with more website conversions than the CRM holds, the diagnosis of LinkedIn conversions higher than the CRM applies this method to click and view credit, duplicate conversions and test submissions.
Apps add more counters to the same problem, and the tutorial on reconciling app install numbers applies this method to store consoles, analytics first opens, a measurement partner and each ad network.
For TikTok, the tutorial on building a monthly TikTok report puts this method into a monthly routine that also separates Spark Ad engagement from the account's organic figures.
Whether explained platform gaps sit inside a wider system of evidence, with one agreed count of conversions and tests on the largest spend lines, is what the marketing measurement audit scorecard rates when a business reviews how it judges its marketing.
Common mistakes
- Looking for the true number. The order system is the complete count; the platform and analytics answer different questions and neither will match it.
- Comparing unlike units. Clicks with users, platform conversions with all-channel analytics conversions, this morning's figure with last week's.
- Reconciling too soon. Platform conversions keep arriving for days after the click; analytics processing has its own schedule.
- Ignoring view-through. The largest single cause on display, video and social, and separable with one platform setting.
- Treating consent loss as a platform error. The gap widens where consent rates fall; check the rate before the tags.
- Changing a window or model without recording it. Every such change moves the numbers and, undated, becomes an unexplained break.
- Fixing the discrepancy by adjusting a report. Tagging faults and duplicate actions are fixed at the source; the rest is documented, not hidden.
- Using the platform's number to decide budget between channels. Each platform claims what it touched; for allocation, lean on tests and on the author's approach described on the paid media analytics expertise page.
