The short answer
The Conversions API sends purchase and other events from your server to Meta, next to the browser pixel. It matters for creative because every test compares cost per result between ads: if events go missing on one path, the wrong ad wins. On our board, 17% of winning ads send traffic to a funnel subdomain (try., offers., get.) rather than the main store, each one a place where tracking can break.
What the Conversions API does
The Conversions API (CAPI) sends conversion events from your server to Meta: purchases, leads, add-to-carts, subscriptions. The browser pixel does the same thing from the visitor's browser. Browsers block, drop and delay a share of pixel events, so Meta recommends a redundant setup: pixel and CAPI together, sending the same events.
Running both creates a second job: making sure each purchase is counted once. Meta deduplicates by event ID and event name. If the pixel sends a purchase with an eventID and the server sends the same purchase with a matching event_id and event_name, Meta keeps one. Deduplication only applies to events received within 48 hours of the first event with that ID.
Meta's Dataset Quality API documentation lists the numbers it uses to grade your setup:
| Metric | What it tells you |
|---|---|
| Event match quality (EMQ) | A score out of 10 for how well your customer data (email, phone, IP, browser IDs) matches events to Meta accounts |
| Event coverage | The share of pixel events also covered by CAPI with shared deduplication keys, as a 7-day average |
| Deduplication keys | The share of pixel and CAPI events sent with each key |
| Data freshness | The delay between when an event happened and when Meta received it |
| Additional conversions reported | How many extra conversions CAPI adds on top of the pixel |
You can see most of these in Events Manager for each event.
Why creative tests depend on it
A creative test is a comparison of reported results: ad A got purchases at $42, ad B at $61, so A wins. Every number in that sentence comes from the events Meta received. Two things can go wrong.
Noise. If a random share of purchases goes missing for every ad, all CPAs look worse and differences take longer to show. You need more spend to reach the same confidence.
Bias. If purchases go missing more on one path than another, the comparison itself is wrong. A simple example: two ads each truly drive 20 purchases on $1,000. Ad A sends people to a quiz on a subdomain where 30% of purchase events never reach Meta. Ad B goes straight to the product page with full coverage. Reported: A at 14 purchases ($71 CPA), B at 20 ($50 CPA). You keep B and kill A, though they sold the same.
And it is not only your report that is fooled. Meta's ranking models learn from the conversion events you send. The GEM model and the models it trains use each person's sequence of actions, including conversions, to predict who will buy. If ad A's buyers are under-reported, delivery learns that ad A converts less and shifts budget away from it. Bad signal makes the algorithm wrong in the same direction as your spreadsheet.
Where winners' traffic actually goes
The bias case is common because winning ads often don't send traffic to a single store domain. In our October 9, 2026 snapshot of 2,443 live Meta ads with $50k+ in estimated spend:
17%
of winners land on a funnel subdomain (try., offers., get., shop., order.)
43
brands send winners to two or more different landing hosts
82
brands send winners to two or more landing page types
184
winners land on a quiz, listicle or advertorial
Leaving out ad-server redirects, the most common subdomain prefixes are try. (109 ads), offers. (64), get. (35), shop. (31) and offer. (27). Many of these are separate funnel pages, often built on a different tool from the store, sometimes with a separate checkout. Each one is a place where the pixel, the server events and the deduplication keys have to be set up again.
The multi-step funnels are also the ones that do well on the board:
| Landing page type | Ads | Share reaching top tier | Median estimated spend |
|---|---|---|---|
| Advertorial | 19 | 15.8% | $316K |
| Listicle | 86 | 14.0% | $232K |
| Quiz | 79 | 13.9% | $319K |
| Product page | 874 | 9.7% | $222K |
| Dedicated landing page | 562 | 9.6% | $220K |
Top tier is the top 10% of the board by Ad Radar's winner score (live 21+ days); the advertorial group is small. These funnels add steps between the click and the purchase, and every step is a chance to lose the event. Brands running them have the most to gain from clean server-side tracking, and the most to lose without it. More on the formats in quiz funnels and listicle landing pages.
One concept, three destinations
Wuffes runs 25 winners across three hosts: its main store, an offer. subdomain and a try. subdomain. The same headline, "Better joints guaranteed in 90 days," sends people to a listicle, a quiz and a product page.
Better joints guaranteed in 90 days
- Est. spend
- $15M
- Days live
- 55
- Format
- Video, 52s
- Variants
- 6
Lands on a listicle on a try. subdomain. Over $14M in estimated spend in 55 days, so the events from that path are carrying a lot of the account's learning.
Better joints guaranteed in 90 days
- Est. spend
- $2M
- Days live
- 71
- Format
- Video, 44s
- Variants
- 3
Same promise, sent to a quiz on an offer. subdomain. If quiz purchases were under-reported, this path would look weaker than it is.
I never thought I would like short press-ons at my age, but here we are.
- Est. spend
- $8.4M
- Days live
- 56
- Format
- Video, 59s
- Variants
- 3
A listicle on a country subdomain. Another separate host whose events need to match the store's.
We can't see anyone's tracking. The point is the structure: when one concept runs to three destinations, comparing them is only fair if all three report purchases equally well.
A signal check before your next test
- Run pixel and CAPI together for every purchase event, on every host that takes orders, including funnel subdomains and third-party checkouts.
- Send the same event ID from both sides. Check the deduplication keys in Events Manager. Missing keys mean double counting or lost events.
- Pass the customer data you legally can. Hashed email and phone, client IP and user agent, and the
fbpandfbcbrowser IDs raise event match quality. - Send server events fast. Deduplication has a 48-hour window, and stale events help neither reporting nor delivery. Watch data freshness.
- Compare Meta's numbers to your store by path. Once a week, compare purchases Meta reports per landing host against your backend orders from Meta traffic. A path that reports much lower than the others is a tracking problem, not a creative one.
- Only then judge creative. Our creative testing framework and creative KPIs assume the events are right. If they aren't, fix that first.
Weak signal also slows the learning phase, because the ad set counts only the events Meta receives. To see which landing page types competitors pair with their winning ads, Ad Radar tags the landing page of every ad and keeps a snapshot of it.
Figures marked as estimated spend come from Ad Radar's model of engagement on public Meta Ad Library ads. They are estimates, labeled as such, and are best used to rank ads against each other.