Broken tracking does not announce itself. It keeps reporting, just wrongly, and every budget decision downstream inherits the error. We find what has quietly died, rebuild it with the fewest moving parts and set a watch so the next break gets caught in days rather than months.
On one of our accounts a checkout signal stopped recording on a single day in May. Nothing errored. No alert fired. Purchases carried on recording perfectly, so revenue and ROAS stayed correct and nobody had reason to look. The signal stayed dead for three months.
What it cost was not money reported wrongly, it was the ability to see. Cost per checkout read a nonsense number, one of the signals the bidding algorithm was told to optimise toward had almost nothing left in it, and a whole layer of funnel diagnosis was gone. We found it during a routine report, dated it to the day and rebuilt it with one less thing to go wrong.
Good Media Buying · Measurement Practice 2026We go through every conversion action you have, count what each one has actually recorded month by month, and find the ones that quietly stopped. Where something has broken we date it, then work back through container history, event names and site releases until we can say exactly what changed.
Browser-only measurement loses conversions to ad blockers, privacy defaults and short cookie lifetimes. Moving measurement server-side recovers a large share of what your ad platforms never see, and gives the algorithms more of the signal they are optimising on. We implement it directly or with partner platforms.
Blended ROAS hides whether you are buying new customers or paying to reach the ones you already had. We separate new-customer acquisition cost from repeat revenue and model the paths that lead to a first order. Where cost of goods is available we report on profit rather than revenue.
Not every conversion should steer spend. We decide deliberately which signals the algorithm bids on and which are there to be read, so a soft or fragile event is never quietly driving your budget. New signals go in as observation first and only get promoted once they have proven themselves.
Conversion rate falling while traffic, CPMs and click-through hold steady is far more often a stock problem than a media one. We cross-reference live spend against real inventory, ad by ad and feed by feed, and flag the money going to product that is sold out or down to a broken size curve.
The whole point of finding a break is never having to find that one again. We put automated checks on the signals that matter, reporting into the channel your team already reads. A healthy run says nothing. A signal that dies says so the same week, not the same quarter.
Every conversion action pulled and counted by month, so a signal that has stopped cannot hide.
Daily data dates the break. Container history, event names and site releases explain it.
Rebuilt with the fewest hops between the site and the platform. Every extra link is a future failure.
Watched firing on the real user journey, then checked again in the data once it is live.
An automated check takes over, so the next break surfaces on its own.
Most tracking work goes wrong in the checking, not the building. A tag that looks right in a config screen, a test run on a page nobody actually arrives at that way, an event that is busy in one tool and absent from another. Each of those produces a confident answer that happens to be false.
So we hold to a few rules, mostly learned the hard way: test the path a real customer takes, not the one that is convenient to test; and never report something as verified when it was only assumed.
A break gets pinned to the day it happened, from daily conversion data. That is what makes the cause findable.
Every system a conversion passes through is somewhere it can silently stop. We remove links rather than repair them.
A new signal goes live as observation only. It cannot move budget until it has proven it counts the right thing.
A failed pull must never render as a real zero. Our checks say "no data" when there is none, instead of quietly reporting nothing happened.
We confirm a tag firing on the journey a real customer takes, then confirm the numbers arrive. Config that looks correct is not evidence.
Book a 30-minute call. We will take a look at what your conversion actions have actually recorded over the last six months and tell you whether any of them have stopped. Most accounts have at least one surprise in them.