Your baseline is a month of your own tracking install

Nobody argues with the principle. Instrument first, then launch. The argument is always about the fortnight it costs, and the fortnight usually wins.

What actually happens

A programme is approved in the second week of a quarter. The creative is ready, the audience is built, and the tracking is described as a technical detail somebody will finish alongside. So the campaign starts, and the tracking gets finished across the following three weeks in the gaps between other work.

Nothing errors during those three weeks. Data arrives. Reports render. Charts draw lines. What the data actually records is a configuration in progress: an event firing twice for a fortnight, a field arriving empty, a conversion counted at the wrong step, a channel labelled direct because the parameters were added late.

Then the quarter ends and somebody builds the trend. The first three weeks are in it, because they are in the system and nobody flags a period as unreliable after the fact. Those weeks become the point every later result is compared against, and every comparison for the next year runs against a number that measured a build rather than a market.

What a proven test event actually is

The phrase gets used loosely, so it is worth being exact. Four conditions, all of them observed by a person rather than inferred from a settings screen.

• A real submission, made through the live page as a visitor would make it, rather than a test fired from inside a tag manager preview.
• Arriving at the named destination, where the destination account was read from the live system in that same session. Not from a screenshot, not from a document, not from anybody's memory of which portal it was.
• With its fields intact. A record that arrives with the email address and nothing else is a record that will fail at routing three weeks later, quietly.
• Visible in the report, not just in the raw store. The gap between an event existing and an event being readable is where reporting disputes tend to live.

Where the same rule applies and nobody applies it

Tracking is the obvious case. The rule is wider than tracking, and the expensive failures are usually in the places nobody thought of as instrumentation at all.

• A form posting into a routing rule. The form reports success. The rule drops the record.
• A webhook into a system that quietly accepts and discards anything with an unexpected field.
• A sending domain that authenticates correctly and delivers to a spam folder.
• A tag moved between accounts, still firing, now populating somebody else's reporting.
• A connector between two systems that syncs in one direction when everybody assumed two.

Every one of those states looks identical to a working state from the marketing side. Nothing errors. Somebody says it is connected, and the word is true and useless at the same time.

When it has already gone wrong

The instinct is to fix it quietly and say nothing, and that instinct is wrong twice over. It leaves the bad period inside the trend, and where data reached a place it should not have, it leaves an unrecorded problem that surfaces later with worse timing.

The honest sequence is short. Stop the flow. Establish exactly what arrived where and over what dates. Mark the affected period as unreliable in the reporting rather than deleting it, so the gap is visible to whoever reads the chart in six months. Tell whoever needs to know on the day. Then re-baseline in the open and say plainly that the line starts again.

None of that is comfortable. All of it is cheaper than the alternative, which is a year of decisions made against a number that measured your own installation.