Campaigns that launched at the start of Q2 are far enough along now that the numbers should be trustworthy. For a meaningful share of the accounts I look at each spring, they aren't, and the reason is never the media buy. It's the tracking setup underneath it, built in a rush before launch and never checked again.
Tagging and tracking are the part of a campaign nobody wants to spend a meeting on. It's not creative, it's not targeting, it doesn't show up in a recap deck as a decision anyone made. So it gets built once, usually by whoever was available the week before launch, and then it runs untouched for months while everyone downstream trusts the numbers it produces. I've opened GA4 configurations in April to find conversion events still firing on a URL structure the site stopped using in February, or a UTM convention that three different team members interpreted three different ways, quietly splitting one channel's performance into what looks like several underperforming ones. In one audit, a "underperforming" channel that was three weeks from being cut turned out to be the account's best performer wearing four different naming conventions. The media team had spent a quarter optimizing around a typo.
The failure mode here is specific, and it's worth understanding why it's uniquely dangerous: bad tracking doesn't usually produce numbers that look obviously wrong. It produces numbers that look plausible and slightly off. A pixel that fires zero conversions gets caught in a day, because zero is loud. A pixel that double-fires on one template, or misses one browser, or drops the sessions that route through a redirect, produces a 15 percent swing, and a 15 percent swing is within the range of things that just happen. Nobody goes looking for a broken instrument when the instrument is telling a story that sounds reasonable. Which means the tracking has to be actively audited on a schedule, not glanced at when suspicion strikes, because the errors that matter are precisely the ones engineered by chance to never trigger suspicion.
There's a compounding cost that makes this worse than a reporting problem. Every platform's bidding algorithm is training on the conversion feed your tracking produces. Feed it data that's quietly wrong for a quarter and you haven't just misreported the past, you've mis-taught the systems buying your future: smart bidding optimizing toward the double-counted event, away from the undercounted one, with total confidence. Fixing the tag fixes the reports immediately. The algorithms take weeks to unlearn what the broken tag taught them, which is why the cost of a tracking error is measured from when it started, not when it was found.
A real tracking audit takes an afternoon, not a sprint, and it has three parts. Fire every conversion event manually and confirm it lands correctly, in the right property, once. Check that UTM parameters are applied consistently across every team touching campaign URLs, not just the ones building this quarter's new creative, because the email team's convention and the paid team's convention diverging is the single most common split I find. Then reconcile platform-reported conversions against your own analytics for a sample week and understand exactly why they don't match, because they never match exactly, and knowing your specific, stable gap matters more than the gap itself. A known 20 percent discrepancy is a calibration constant. An unexamined one is a landmine with a QBR date on it.
Then put the afternoon on the calendar quarterly, tied to something that already recurs so it can't be skipped, and assign it a name. Tracking decays the way the site changes: every redesign, every new landing page template, every consent banner update is a chance for an event to silently stop firing, and none of those projects will think to tell the analytics owner, because in most organizations there isn't one. Do this now, six weeks into the quarter, while there's still time to correct the number the rest of your reporting depends on. The alternative is finding the broken pixel in the QBR, after three months of decisions were made on data that was quietly wrong the entire time, and after the algorithms spent those three months learning it by heart.