THE MACHINERY OF THE TRUSTED NUMBER

What has to be true before a number is allowed to move a decision

A dashboard says the channel produced three hundred and twelve customers last month.

Everyone in the company treats that as a fact about the world, the way they would treat the temperature or the date. It is on a screen. It has a decimal place. It was generated by a system nobody in the room built, and it arrived without a human hand visibly touching it, and those two properties together make it feel like something found rather than something made.

It was made. Every number on every dashboard was made, by a chain of decisions that a person made, most of which nobody in the room could name.

Somebody decided what counts as a customer. Somebody decided which of the four things a person touched before buying should receive the credit. Somebody decided how long after a click a purchase still counts. Somebody wrote a rule for what happens when the same human appears twice on two devices. Somebody decided whether refunds get subtracted, and when.

Change any one of those decisions and the number changes. Not slightly. It can double.

The three hundred and twelve is not a measurement of reality. It is the output of a machine, and the machine has opinions.

PART ONE: EVERY NUMBER IS A CHAIN, AND THE CHAIN IS AS STRONG AS ITS WORST LINK

Between the thing that happened in the world and the figure on the screen sits a chain, and it is longer than almost anyone thinks.

Something occurred. A person did something. Then that something had to be captured, by a tag or a script or a sensor that had to be present, correct, and functioning at that exact moment. Then it had to be attributed to a cause, by a rule somebody wrote. Then it had to be deduplicated against everything else, by another rule. Then it had to be aggregated, and the aggregation had to choose a window. Then it had to be displayed, which required choosing what to compare it against.

Six links, minimum. A break at any one of them and the number is wrong, and here is the property that makes this dangerous rather than merely annoying:

A broken number looks exactly like a working one.

It has the same font, the same decimal places, the same position on the same dashboard. There is no visual difference whatsoever between a figure produced by an intact chain and a figure produced by a chain that silently snapped three weeks ago. Nothing turns red. Nothing alerts. The number is simply, quietly, wrong, and it goes on being wrong, and decisions go on being made against it.

Silence Is Not Evidence of Health

The reason bad numbers survive is that nothing complains.

An analytics tag that stops firing does not raise an error. It simply reports zero, and zero is a legal value, and a metric that has dropped to zero looks like a business problem, so the meeting is about the business problem, and the tag is never examined.

Worse, an attribution rule that overcounts does not report an error either. It reports success. It reports more customers, more conversions, better performance, and nobody investigates a number that is bringing good news, because investigating good news feels like a personality defect.

Which produces the asymmetry that governs this entire field: errors that flatter you survive longer than errors that hurt you, and they are the same kind of error.

For a period in the late 2010s, Facebook disclosed a series of measurement errors in its advertising platform, and the pattern in them is the whole lesson. The average time a video was watched was overstated, in some accounts by a very large factor, because the calculation divided total watch time by the number of people who watched for more than three seconds rather than by everyone who saw it. Organic reach was overstated. Referral traffic was misreported. Time spent was miscounted. The errors were real, and the corrections were made, and none of that is the interesting part. The interesting part is that these numbers had been sitting on screens for years, in front of thousands of sophisticated advertisers whose full time job was to read them, and had been used to move budgets worth billions of dollars, and nobody caught it. Not because the advertisers were stupid. Because the errors made the platform look good, and a number that makes your channel look good does not generate the small itch of suspicion that a number making it look bad generates instantly. Nobody audits the metric that is bringing them what they hoped for. That is not a fact about Facebook. It is a fact about people, and it applies to every dashboard in every company, including the one you are looking at.

PART TWO: THE FOUR PLACES A NUMBER BECOMES A LIE

The Definition

What counts as a customer. What counts as a lead. What counts as active.

These sound like clerical questions and they are the deepest ones, because two departments using the same word for different definitions will generate two different numbers, and both will be correct, and the resulting argument will never be resolved, because it is not an argument about the world. It is an argument about a word, being conducted as though it were about the world.

The tell is unmistakable once you know it: two teams, two numbers, both confident, neither wrong. That is never a data problem. That is always a definition problem.

The Capture

Nothing gets counted that was not recorded, and recording is a physical event that can fail.

The tag was not on the checkout page. The tracking broke during the redesign and nobody noticed for eleven days. Ad blockers ate a fraction of the events, and that fraction is not random, it correlates with exactly the technical, affluent audience the business cares most about. The mobile app version fires different events than the web.

The data is not a mirror of reality. It is a mirror of what the instruments happened to catch, and the instruments have blind spots that are shaped like specific kinds of customers.

The Attribution

A person saw an ad on Monday, searched on Wednesday, clicked an email on Friday, and bought on Saturday.

Which of those gets the customer?

There is no fact of the matter. There is only a rule, and the rule was chosen by someone, and every choice produces a different picture of which channels are working. Give the credit to the last click and email looks brilliant. Give it to the first touch and the ad looks brilliant. Split it evenly and everything looks mediocre.

The channels did not change. The rule changed. And most companies have never once looked at the rule that is generating the numbers they use to allocate their entire budget.

The Window

Count purchases within seven days of the click and one number appears. Count within thirty days and a bigger one does. Count within ninety and bigger still.

None of these is wrong. But comparing a channel measured on a thirty-day window to a channel measured on a seven-day window is not a comparison of two channels. It is a comparison of two windows, and the outcome was determined before either channel did anything.

PART THREE: THE PRECISION ILLUSION

A number that reads 312 feels more true than a number that reads about 300.

This is a bug in the observer, and it is exploitable, and it is exploited constantly, including by well-meaning systems that have no intention of deceiving anyone.

Precision is a property of the display. Accuracy is a property of the relationship between the number and the world. They are entirely independent, and the human eye reads the first as evidence of the second.

A figure carried to two decimal places, produced by a chain of six shaky links, is not more trustworthy than a rounded estimate. It is less trustworthy, because it is wearing a costume of certainty that its own construction cannot support, and that costume suppresses exactly the suspicion that should have been raised.

On the twenty-eighth of February, 1991, an Iraqi Scud missile struck an American army barracks in Dhahran, killing twenty-eight soldiers. The Patriot battery that should have intercepted it did not fire. The cause was a number. The system tracked time in tenths of a second, stored in a binary register that could not represent one tenth exactly. The tiny error, invisible at any single moment, accumulated as the system ran. After a hundred hours of continuous operation, the internal clock was off by about a third of a second. A third of a second. In that time, a Scud travels roughly half a kilometre. The radar looked for the incoming missile in a range gate that was, by then, in the wrong place, found nothing there, and concluded that the detection was spurious. Every number in that system was displayed to full precision, the whole time. Nothing on any screen ever indicated that the clock had drifted. The system was confidently, precisely, and lethally wrong, and it had been getting wronger by the hour, and the display gave no hint at all. Precision is not accuracy. The display will never tell you the difference. It is not built to.

PART FOUR: WHAT MAKES A NUMBER TRUSTABLE

Trust is not a feeling about a number. It is a set of conditions that either hold or do not, and each one can be checked.

It reconciles against something with money in it. The count of customers should be forced to agree with the count of transactions in the payment processor, which is the one system in the business that cannot lie for long, because it is where the actual money is. Any number that has never been reconciled against a financial record is a number nobody has checked.

Its definition is written down, in one place, and shared. Not held in someone’s head. Written, so that when two people say customer they are provably saying the same thing.

It has a known blind spot. Every instrument has one. A person who can state the blind spot of their own number is a person who understands their number. A person who says the data is complete has simply not found the hole yet.

It survives a change in method. Compute it a second way, crudely, from a different source. If the two disagree by ten percent, you have learned something. If they disagree by four hundred percent, one of them is not measuring what its name says it measures.

It has been wrong before, and someone noticed. A number that has never been caught being wrong is not a number with a good track record. It is a number with no track record, because nothing in the system was ever capable of catching it.

PART FIVE: WHAT CHANGES IN THE PERSON WHO SEES IT

The dashboard stops being furniture.

A number appears, and instead of the number, the chain behind it appears, and the mind moves along it, quickly, almost without effort. What is being counted. What caught it. Which rule gave it the credit. Over what window.

This looks like skepticism from the outside, and it is not. Skepticism is a stance and it is exhausting and it makes a person tiresome in meetings. This is closer to depth perception. The number stops being a flat object on a screen and becomes a thing with a construction behind it, and once that construction is visible, the appropriate amount of confidence in the number becomes obvious, and it is usually not the amount that everyone in the room is currently applying.

And then something practical follows, which is the point of all of it.

The person stops arguing about what the number means and starts asking how it was made. Those two conversations look similar and they go to completely different places. The first one can run for a year. The second one ends in about twenty minutes, in a definition document or a broken tag, and then the argument is simply over, because it was never really an argument at all.

SYNTHESIS

THE CHAIN BEHIND EVERY NUMBER ON EVERY DASHBOARD

  the thing happened
        ▼
  CAPTURE      a tag, a script, a sensor. it can be absent.
        ▼
  DEFINITION   what counts as a customer. someone decided.
        ▼
  ATTRIBUTION  which touch gets the credit. a rule, not a fact.
        ▼
  WINDOW       7 days, 30, 90. the answer changes with the choice.
        ▼
  AGGREGATION  averaged over what, compared against what
        ▼
  312 customers          ◄── arrives looking like a fact about the world


  A BROKEN NUMBER LOOKS EXACTLY LIKE A WORKING ONE.
  SAME FONT. SAME DECIMALS. NOTHING TURNS RED.


  AND THE ASYMMETRY THAT DECIDES EVERYTHING:

    an error that HURTS you    ──►  gets investigated within a day
    an error that FLATTERS you ──►  becomes the number you plan around

    they are the same kind of error.

Every number was made, not found. Between the event in the world and the figure on the screen sits a chain of six decisions that a person made, and changing any one of them changes the number, sometimes by a factor of two.

A broken link in that chain produces no error, no alert, and no visual difference. The number simply becomes wrong and stays wrong, and the errors that flatter you survive longest, because nobody has ever audited a metric that was bringing them good news.

Four places the number becomes a lie: the definition of the thing counted, the instrument that captured it, the rule that assigned credit, and the window over which it was summed. Two teams with two different numbers, both confident, is never a data problem. It is a definition problem, being fought as though it were about the world.

And the display lies in a way that is almost impossible to resist. Precision is a property of the screen. Accuracy is a property of the relationship to reality. The eye reads the first as evidence of the second, and it is not, and no dashboard ever built will tell you the difference.

A number earns trust by reconciling against money, by having its definition written down, by knowing its own blind spot, by surviving a second method, and by having been caught being wrong at least once by something in the system that was capable of catching it.

Everything else is faith in a screen.

CITATIONS

United States General Accounting Office (1992). Patriot Missile Defense: Software Problem Led to System Failure at Dhahran, Saudi Arabia. GAO/IMTEC-92-26. A cumulative floating-point error in the tracking clock caused the range gate to be misplaced after long operation. Every display remained fully precise throughout.

Barreto, T. (2016). Facebook’s video metric error and the advertiser response. Reported in the Wall Street Journal, September 2016. A sequence of disclosed measurement errors in a major advertising platform, most of which overstated performance and none of which were caught by the advertisers reading them daily.

Lewis, R.A., & Rao, J.M. (2015). The unfavorable economics of measuring the returns to advertising. Quarterly Journal of Economics, 130(4), 1941-1973. Observational advertising metrics routinely fail to isolate causal effect, and the statistical power needed to do so is far greater than practitioners assume.

Kohavi, R., Tang, D., & Xu, Y. (2020). Trustworthy Online Controlled Experiments. Cambridge University Press. The trust checks used by large-scale experimentation platforms: sample ratio mismatch, guardrail metrics, and reconciliation against independent sources. Numbers earn trust by being tested, not by being displayed.

Meng, X.-L. (2018). Statistical paradises and paradoxes in big data. Annals of Applied Statistics, 12(2), 685-726. A large sample with a small systematic bias in who gets captured is far less trustworthy than a small unbiased one, and the size of the dataset actively disguises this.

THE MACHINERY OF THE GOAL BEFORE THE SPEND. The number that gets read against the goal has to be believable first, which is what this is about.
THE MACHINERY OF THE RATIO NOT THE TOTAL. Once the number can be trusted, the next question is what it was divided by.
THE MACHINERY OF THE COST OF A CUSTOMER. The number most often trusted without ever being reconciled against the money.
THE MACHINERY OF SIGNAL AND NOISE. The deeper law. Nothing is intrinsically signal, and the question decides what the data is allowed to say.