First, confirm the duplication
Not every gap between Meta and the shop is double counting. Meta credits conversions to the date of the ad interaction under the account's AttributionAssigning conversion credit to touchpoints. setting, while the shop counts orders on the day they happen, so short or open periods disagree for ordinary reasons. Pick a closed week that ended a while ago and compare unattributed event counts: the number of purchase events Events Manager received against the number of orders the shop recorded for the same days. Received events should sit close to orders, a little below because of consent and blocking. If received events clearly exceed orders, you have duplication; if attributed results in Ads Manager exceed orders but received events do not, the question is attribution, covered in the measurement hub.
Next, open the affected event's detail. Note which connection methods are listed, how many copies each sends, and the deduplication share. Then open a single test order in Test Events and look at both copies side by side: name, event ID, timestamp. That one order usually tells you which row of the table applies.
Diagnosis table
Diagnosis
| Symptom | Check | Likely cause | Fix |
|---|---|---|---|
| Browser and server copies both arrive, deduplication share near zero | Event ID on each copy of one test order | No shared event ID, or each side generates its own | Generate one ID per action and send it on both copies |
| IDs match, yet both copies are still counted | Event name on each copy | Event names differ between browser and server | Use the same standard event name on both sides |
| Two browser requests per action before any server event | Pixel Helper on the confirmation page | Pixel installed twice, hard-coded plus a plugin or tag manager | Remove one install and keep the one you can inspect |
| Duplicates cluster in batches or appear a day later | Event time against received time for server copies | Server event sent long after the browser event | Send server events in near real time with the original event time |
| Server count alone exceeds orders | Connections listed per dataset; partner integrations | A partner integration and a custom server setup both sending | Keep one server sender per event and switch off the other |
Fix, verify and prevent each cause
No shared event ID
Without a common ID, Meta has no reliable way to know that the pixel purchase and the server purchase are the same order. The usual faults are a pixel tag with no event ID parameter at all, or each side inventing its own value.
Fix: derive the ID from the record that represents the action, such as the order number, expose it to the browser through the data layer, and have the server read it from the same record.
Verify: in Test Events both copies of a test order show the identical ID and the detail view reports them as deduplicated.
Prevent: write the ID rule into the integration's specification; the Conversions API guide sets out the requirements that make deduplication work.
Event names that differ
A partner app that sends a custom name, or a server payload that capitalises differently from the pixel, produces two events that Meta treats as unrelated even when the IDs agree.
Fix: map both sides to the same standard event name.
Verify: the dataset overview lists one purchase event with two connection methods rather than two separate events.
Prevent: keep a short naming list beside the tag container and the server code, and check the dataset overview after any integration update, as the Conversions API guide recommends for verification.
Pixel installed twice
This duplicates on the browser side before any server event exists. A theme snippet left behind after a move to Tag Manager, or a commerce platform's built-in pixel alongside a manual tag, fires every event twice.
Fix: remove one install, keeping the one you can test and version.
Verify: Pixel Helper shows one pixel and one event per action on each page.
Prevent: record where the pixel is installed and who may change it; the troubleshooting table in the Meta Pixel setup tutorial lists a doubled pixel as the first thing to rule out.
Server copy arriving too late
Meta documents a deduplication window. A server event uploaded in a nightly or weekly batch, or retried much later after an outage, may fall outside it and be counted as a new action.
Fix: send web events from the server as the action happens, with the original event time, and make retries reuse the original ID.
Verify: the diagnostics tab no longer warns about delayed events, and received time sits close to event time.
Prevent: monitor the integration's queue and alert when sends back up; the Conversions API guide treats timing as part of the deduplication contract.
Two server senders
A store platform's partner integration often switches on server events by default. If a developer then builds a custom server integration, or a server-side tagging container, the same order reaches Meta twice from the server with different IDs.
Fix: choose one sender per event and disable the other in its own settings.
Verify: the dataset lists one server source for the event, and the server count over a closed week sits at or below orders.
Prevent: keep a register of every system allowed to send to the dataset, described as part of the setup options in the Conversions API guide.
Prevention: one owner per event source
Duplication is an organisational failure more often than a technical one. The pixel was installed by an agency, the platform integration was enabled by the store manager, the server container was built by a contractor, and nobody holds the full picture. The cure is a single owner for the dataset and a short written map: each event, each sender, the field that supplies the event ID, and the event name. Any new sender is added to the map before it goes live, not after. Where several platforms read the same purchase, the tutorial on ad platform tags from one data layer turns that map into a mapping table the container is built from.
Build three checks into that ownership. After any release that touches checkout or tags, place a test order and confirm in Test Events that it arrives once from each route with matching name and ID. Every week, compare received purchase events with shop orders for a closed period and investigate when received events exceed orders. Every month, read the diagnostics tab and the deduplication share for each key event and record them, so a quiet decline is noticed. Recompute return on ad spend (ROASRevenue generated per unit of ad spend.) with the ROAS calculator from shop revenue rather than reported value whenever the counts disagree, so decisions are not made on inflated figures while a fix is pending. The Conversions APIServer-side event sharing with Meta. is only as trustworthy as the weakest of those checks.
A skincare brand that counted every order twice
A skincare brand enables its store platform's Conversions API integration in March. Over the next closed week Events Manager receives 1,180 purchase events against 602 orders, and reported ROAS rises from 2.4 to 4.5 overnight with no change in spend. The event detail shows browser and server copies with a deduplication share of 3%. A test order reveals the cause: the pixel, installed months earlier through Tag Manager, sends an event ID built from a timestamp, while the platform integration uses the order number. The team changes the Tag Manager variable to read the order number from the data layer. The following closed week records 611 received events for 598 orders, a deduplication share of 91%, and a reported ROAS of 2.5.
Common mistakes
Comparing attributed results with orders and calling the gap duplication. Switching off the browser event to make the numbers match. Fixing the ID on purchases and forgetting leads or add-to-cart events that use the same tags. Leaving a test event code in production payloads. Assuming deduplication works because the setup wizard finished; if events do not arrive at all, start from the pixel events not showing page instead, and once counts are right, the low event match quality page covers how well those events match people.
