Signal vs noise: what actually belongs on a journey map

You have seen the map. Somewhere around step four there is a number in a coloured box telling you that 34% of customers reported frustration at this point in the journey, and it is a good number, the kind that makes an artefact feel like it was built on evidence rather than workshop opinion. Then you open the same map a week later and the number is identical, and a month after that it still has not shifted by a decimal place.

The temptation is to conclude that the metric is stale, but that is not quite what has happened. It's not that the number did not fail to move; it's just that nobody looked at it again. It was captured once during discovery and has been sitting there ever since, doing the job it was actually hired to do, which was to make the map look rigorous in a steering committee. Which means the interesting question is not why the metric went stale, but what anyone was going to do when it moved, and whether that was ever decided in the first place.

What is signal and what is noise

The word noise misleads people, because it sounds like a synonym for bad data. It is not. Noise is real, correctly measured variation that simply carries no instruction, and every operational metric you own produces it constantly. Your abandonment rate at a given step will wobble between 21% and 26% more or less forever, for reasons nobody can name and nobody needs to, and none of that movement means anything at all. Signal, by contrast, is movement that is unlikely to be routine and that would change what somebody does. Both halves of that definition matter, because a statistically genuine shift in a metric that nobody owns and nobody can influence is still noise in every sense that affects your work.

The consequence is uncomfortable for anyone who has put a figure on a map. Suppose abandonment at a step moves from 22% to 25%. Is that a signal? You cannot answer, and no amount of further thinking or stakeholder discussion will get you there, because the answer depends entirely on what the previous twenty readings looked like. If they have been bouncing between 19% and 27%, then 25% is an ordinary Tuesday and you should ignore it. If they have sat tightly between 21% and 23% for six months, then something has changed and you should go and find out what. The figure is identical in both cases and the conclusion is opposite, which tells you that a number without its own history is not evidence, it is decoration.

None of this is new statistically - Deming's separation of common cause variation, meaning the system behaving as it always has, from special cause variation, meaning something genuinely new, is the same idea dressed in older language. You do not need to teach anyone control charts to make use of it, you just need to stop treating a bare figure as though it says something on its own.

Which signals matter more than others

There are two filters worth applying here, and practitioners collapse them into one far too often, which is how you end up with a map full of numbers that are technically defensible and practically inert.

The first filter asks whether the movement is real, and it starts with volume and cadence. Does the metric arrive often enough, and in enough quantity, that you could distinguish movement from wobble even in principle? This is where I will lose a portion of the audience, because the honest answer for most survey data on most journey maps is no. A quarterly relationship survey sliced down to a single step of a single journey takes a few hundred responses and turns them into a few dozen, with confidence intervals wide enough to drive a truck through, and what you are left with is not a measure of how customers experienced that step so much as a record of who happened to respond. Meanwhile the telemetry that could carry the analysis properly, things like abandonment, task completion time, repeat contact rate and call volumes by reason code, arrives daily in volume and usually sits in a system that was never invited to the mapping workshop. It is also worth confirming that the metric's definition has held over the period you are looking at, since a good proportion of dramatic shifts in enterprise reporting turn out to be someone quietly changing how a field is calculated.

The second filter asks whether the movement warrants a response, which is a different question entirely. Is the metric attached to an outcome somebody is accountable for, such as cost to serve, churn, completion or repeat contact? Is there a lever, given that a signal sitting at a step nobody can influence is interesting rather than important? And is there an owner and a threshold agreed ahead of time, because if you cannot say what number would trigger what action, you are not measuring anything, you are furnishing the room. As a rough ordering when you have to choose, favour leading over lagging, operational over attitudinal, step level over journey level, and owned over orphaned. Nobody will ever use a weighted scoring model, but people will use four rules.

How to present signals on a journey map

This is the part that tends to be unpopular, so I will put it plainly: a journey map is a coordinate system, not a dashboard. Its real job is to give an organisation a shared vocabulary for where things happen, so that finance, operations and the contact centre can argue productively about the same step rather than talking past each other. Live measurement belongs in a system built for measurement and keyed back to that map. Asking a poster to carry live data is asking the wrong tool to do the work, and the staleness you noticed at the start was never an accident of neglect; it was structurally guaranteed from the moment the number went on. That being said, digital live versions of journey maps can still have live data hooked in, but again it is a coordinate system and not a dashboard because the layout serves a different purpose to what dashboards are designed to do.

Where numbers do appear on a map, a few rules help. Keep it to one metric per step, since two usually signals that the step has not been understood well enough to choose between them. Never show a bare figure on its own, because direction and recent range carry the meaning, and a spark line the size of a fingernail communicates more than a percentage with two decimal places. Annotate what changed rather than only what the value is, since a note recording that the new IVR went live in March explains more than any amount of precision. Mark provenance and date on everything, so a reader knows what was measured, from what source and when. And keep the baselines visually distinct from the live measures, because a baseline holding steady is doing exactly what it should, whereas a live metric that has not moved in a month means nobody has looked.

Can a signal turn back into noise?

It can, and this is the failure that almost nobody plans for, largely because it happens slowly. Sometimes the problem simply got fixed, meaning the signal did its job and should be retired, though very few organisations ever retire a metric, which is precisely how dashboards die of overcrowding. Sometimes the lever changed underneath it, so a process redesign leaves the metric measuring something other than what it was chosen for while the numbers keep arriving as though nothing happened. Sometimes attention distorts it, where reporting a measure upward bends behaviour toward the measure until the movement stops reflecting customers at all. And sometimes the map itself has drifted, so the step the metric assumes no longer exists in that shape.

The remedy is unglamorous, which is a review cadence and a willingness to let things expire. Quarterly is defensible for most organisations. Put three questions to every metric on the map: has it moved, did anyone act on it, and would anybody notice if it disappeared tomorrow. Anything that fails all three comes off, and the map is better for the space.

Where to start

Take the map you have now and, for every number on it, establish when it was last measured and who owns it. The answer will be uncomfortable, and that discomfort is the finding rather than an obstacle to it. Then pick a single step and a single operational metric and plot its last twenty observations before you conclude anything at all about its current value. Agree a threshold and a response before the next reading rather than after it, and retire whatever fails the three questions above.

You will end up with a map carrying fewer numbers, which tends to feel like a loss until you notice that the remaining ones are attached to decisions. The stale figure was never really the problem. The absence of anything that was going to happen because of it was.