The Number Was Wrong for Six Weeks and Nobody Noticed

Here is a shape you have probably lived some version of. A quarterly review opens with a chart that has a cliff in it. Organic sessions, steady for a year, then a drop that never comes back.

The room does what rooms do. One person blames an algorithm update. Another says the category is softening. A third says this is exactly why we need to be serious about AI search. Somebody writes “audit content strategy” on a whiteboard and everyone feels productive. Nobody in the room says the thing that turns out to be true, which is that broken measurement does not throw an error. It renders a clean chart and lets you present it.

The cause was a website deploy. A contractor cleaning up the site header removed a script tag nobody could identify, because nobody could identify it. It was the analytics tag. For six weeks the site was fine, the traffic was fine, and the measurement was gone.

Two monthly reports went out on that data. Nobody caught it, because nothing ever looked broken. The chart rendered perfectly. It just rendered a lie.

Charts do not know when they are empty

Here is what makes this failure mode so durable: broken measurement does not announce itself. It throws no red banner. It pages nobody at 3 a.m. the way a down server would. A chart takes whatever rows it is handed and draws them, and a chart with no rows looks exactly like a chart with bad news.

Your engineering team has monitoring for this. If the site goes down, half the company gets a text before anyone finishes a sentence. Marketing measurement gets none of that. The pipes that carry your numbers (tags, pixels, webhooks, automation tasks, CRM field mappings, ad platform connections) are load-bearing infrastructure held together by naming conventions and one person’s memory, and they are built to fail quietly.

Zero is a legitimate value. That is the entire problem. “Nothing happened” and “nothing got recorded” produce the same pixel.

Nobody files a bug report on a number that went up

The six-week blackout is the embarrassing version. The version that actually costs money runs the other direction.

Somebody adds a second conversion event for the same form fill because the original had a confusing name and nobody wanted to touch it. Now every submission counts twice. Lead volume climbs. The paid channel that owns most of those forms suddenly looks like the strongest line in the budget. So you shift money into it. You plan headcount against it. You put it in the deck as the thing that is working.

Nobody investigates good news. There is no incentive to. When a number falls, three people have to explain themselves, so eventually somebody digs. When a number rises, every person near it has a reason to accept it quietly and take partial credit. Flattering bad data can live inside a company for a year without being challenged once.

The scariest number in your reporting is not the one that fell off a cliff. It is the one that quietly went up in a month when nobody remembers doing anything different.

This is not a tooling failure. It is an ownership failure.

The instinct is to go buy something that watches the pipes. Sometimes that helps. But the reason six weeks passed is not that a piece of software was missing from the stack. It is that nobody in the building had the job of asking “is this real?” before the number went into a slide.

Everyone who touched that chart assumed somebody upstream had checked it. The analyst assumed the platform was right. The manager assumed the analyst had looked. The VP assumed the manager would flag anything weird. The CEO assumed all three had done their jobs. Diffused responsibility is not a data problem, but it manufactures one every quarter.

And checking is genuinely unpleasant. Verifying one suspicious number means logging into four systems, exporting three files, reconciling two date ranges that do not agree on what a week is, and losing a morning to confirm something you already suspected. So people do it once, find nothing, and quietly stop doing it. The price of asking is high enough that most companies only pay it after the number becomes embarrassing in public.

What actually works is boring

• Keep a changelog nobody reads until they need it. One shared doc, one line per event, always dated: deploys, tag changes, CRM field renames, form vendor swaps, campaign launches, automation edits, permission changes. Most mystery movements resolve about as fast as you can scroll a doc, once you line up the date of the drop against the date of the change. Teams that cannot do that theorize instead, and theories are free, which is why they multiply.
• Know the shape of a normal week. Not the target. The shape. If you cannot say roughly how many demo requests a normal Tuesday produces, you cannot tell abnormal from broken, and you will accept whatever the screen says.
• Treat every sharp move as a bug until proven a trend. Real shifts in demand are usually gradual and show up in more than one place at once. Instrumentation failures are sharp, precisely dated, and isolated to a single source. If revenue is flat and one channel collapsed, go look at the pipe before you go look at the market.
• Make checking cheap. This is the one that actually changes behavior. Nobody skips verification because they do not care about accuracy. They skip it because it costs a morning. When answering “does this match what the CRM thinks?” takes two minutes instead of half a day, people start doing it casually, and casual checking is what catches things early.

The uncomfortable part

There is a number in your reporting right now that is wrong. Not wrong by a rounding error. Wrong because a pipe broke, or a field got renamed, or a script got deleted by somebody doing exactly what you asked them to do. It is being read aloud in meetings. It is probably in a plan for next year.

You are not going to find it by staring harder at the same chart. Nothing on the screen is going to look suspicious, because nothing on the screen ever does. You find it by making it cheap enough to ask a second question, and then a third, until the story either holds together or falls apart in front of you.

That is a large part of why we built THE DASHBOARD at I Hate Marketing: put the whole stack in one place so that checking one number against the system sitting next to it is a question you ask in passing instead of a project you have to schedule. One flat price, no seat math, no token meter. If you want a look, it is at ihatemarketing.ai. And if you do not, go start the changelog anyway. That one is free.