Conversion tracking setup: one count per real event
There are two kinds of conversion tracking setup. The first passes the launch test: the tag fires, the test purchase appears, everyone moves on. The second is still telling the truth six months later, after a theme update, a consent banner, an app install, and two well-meaning edits by different people.
Almost everyone builds the first kind. We know because we audit the results at scale. Across 936 live sites our scanner examined recently: one in five had duplicate tracking corrupting its counts. 28% had consent configurations dropping conversion data. 37% had no working analytics at all. These are not exotic failures on neglected sites; they are the steady-state of setups that passed a launch test once and were never checked again.
Already have a setup? A free Mimetic scan checks it against everything on this page from the outside, the way a visitor's browser experiences it, and ships corrections as pull requests you review. Starting from zero? It provisions Google Analytics and Search Console you own, one click.
The doctrine is one sentence
One real-world outcome, one event, one primary conversion, one count. Every failure mode we find at scale is a violation of one clause of that sentence.
- Two tag layers (a plugin and a pasted snippet and a Tag Manager container) is how one purchase becomes two counts. One in five sites we scan is in this state or flirting with it, and among sites with GA4, one in five runs two properties at once.
- Two conversion actions counting one outcome (a web event and an offline upload, both primary) is the same sin at the ad-platform layer, and it steers automated bidding with doubled money.
- Zero counts for a real outcome (the event renamed in a refactor, the key event pointing at the old name, the primary action that has never fired) is the inverse violation, and the more expensive one, because bidding algorithms trust the zero.
Setup, then, is mostly subtraction: pick one tag layer, one source of truth for events, one primary per outcome per platform, and delete or demote everything else. The rebuild instructions for any specific stack are a solved problem; your developer or an AI assistant can produce them on demand. What neither will volunteer is the discipline: every additional layer, tool, and "backup" tag you keep is a future double-count with a launch date you did not choose.
The two design decisions that cannot be retrofitted
Deduplication keys. The day you add a second source for the same event (a server webhook backing up the browser event, a CRM upload joining a web action), the two must share an ID so platforms can match them. Retrofit this after months of unkeyed dual-sending and the historical data is simply corrupt. Send transaction_id from day one, everywhere, even while you have only one source. It costs nothing and it is the difference between "we added server-side tracking" and "our revenue has been 40% overstated since March," a sentence we have had to say to real people.
Consent as configuration, not decoration. 29% of sites we audit fire trackers before any consent decision; 28% drop conversions through consent misconfiguration. Both directions are losing moves: the first is legal exposure in opt-in markets, the second is voluntarily deleting your own data in opt-out markets that never demanded it. Consent defaults deserve the same review rigor as pricing, decided per region, on purpose.
Verification is a ratio, not a checkmark
Launch-day verification everyone does: the event fires, the test purchase shows up. The verification that matters is the one almost nobody institutionalizes: the ratio between independent counts of the same reality. Platform orders vs analytics purchases vs ad-platform conversions, same date range, monthly. The three never match exactly (attribution differs by design), but their ratios should hold steady. A moved ratio is a break that shipped, caught in five minutes, months before it would otherwise surface.
And alarm on zero. A daily-firing conversion that reads zero for a day is a page, not a curiosity. Every failure statistic in this article is a count of breaks that stayed silent for months because nothing and no one was watching for absence.
Mimetic institutionalizes exactly this: continuous verification that your tracking still measures what your business does, one count per real event, with every fix shipped as a pull request you review and merge. Related: GA4 not tracking conversions, server-side tracking.