Server-side tracking: see the orders your pixel misses.

Client-side tracking was built for a browser that cooperated. That browser is gone.

Ad blockers strip the pixel before it loads. Safari's Intelligent Tracking Prevention caps cookies written by scripts at 7 days, so a shopper who returns next week looks new. Since iOS 14.5, Apple asks every app user whether they allow tracking; Adjust puts the industry average opt-in at 35% in 2025, so two in three iPhone users send degraded data. And a shopper who taps Buy and closes the tab may have paid on the server while the pixel never fired.

Trackingplan's 2026 report finds that server-side implementations recover up to 30% of previously missed conversion data. The exact share depends on your market and your audience's devices. Our Meta ads report found a different distortion in the other direction: for Bulgarian stores, Meta reported 2.3 times the orders the store itself could see (white paper). Either way, the ad platform is not seeing what the store sees.

It is a learning problem, not a reporting problem

Meta's and Google's algorithms learn from the purchase events you send and look for more people like the buyers. When the missing buyers are mostly Safari users, iPhone users and ad-blocker users, the algorithm learns from the easiest-to-track customers and optimizes to find more of them. Fixing the reporting fixes what the algorithm learns.

The browser still visits your site. The difference is who reports. With a pixel, the visitor's browser calls Meta. With server-side tracking, your server calls Meta's Conversions API after the order is written to your database.

Three details make it work. Your server captures the click id Meta adds to every ad click (the fbclid in the URL) and stores it with the session. When the shopper buys, five minutes or five days later, the server sends the purchase with that click id, the hashed email and the order value. And you keep sending the pixel in parallel, deduplicated by event id, so nothing is counted twice.

Meta scores how well your events match known people as Event Match Quality. Click id plus hashed email is what lifts that score; the score is what decides how much of your spend the algorithm can learn from.

Consent still applies

Server-side tracking does not remove the need for consent under GDPR or similar laws. It changes where consent is enforced: on your server, where you decide what is sent to whom, rather than in a banner that a script may or may not honor. If a shopper withdraws consent, you stop sending their events at the source.

Most guides stop here: set up the Conversions API, recover the missing orders, watch the reported return improve. That is real. It is also the smaller half.

Two customers each spend the same amount today. One never returns. The other buys four more times this year. A purchase event, pixel or server, sends the same number for both, so the algorithm treats them as identical and looks for more of either.

The agent sends two signals: the real purchase, and the predicted value of that customer. Meta and Google can then optimize for people who look like the second customer. Purchases are counted once, and your existing customers are excluded from acquisition, so you stop paying to re-acquire people you already have (ads audiences).

One womenswear brand on Shopify ran it: Meta audiences built from predicted value, with clean server-side purchase events. The daily budget went up 2.1 times over 79 days and the return held within 8%, with 88% more store revenue per click, same 60 days year on year (case study).

Weeks 1 to 2: the basic connection

Capture the click id, store it with the session, send purchases with click id and hashed email, run pixel and server in parallel with deduplication. On Shopify and the other plug-in platforms the agent does this on the day it goes live.

Weeks 3 to 6: validation

Compare pixel events with server events and find the gap. Add match keys (phone, your own customer id). Watch the match quality score climb, then reduce the pixel's role.

Weeks 4 to 8: value signals

Turn on predicted customer value next to the purchase event, exclude existing customers from acquisition, and give the algorithm two to four weeks to learn. Then judge the change by orders in your store, not by the platform's report.