Managing queues: what you have to measure - and what a door counter cannot see
At a German theme park, the share of visitors who gave up queuing for the most popular ride fell from 13.5 to 6.5 percent. Nobody was hired, the ride did not get faster, and the queue did not get shorter. The only thing that changed was a sign showing how long the wait actually was.
That is the whole lever, and it is also the whole catch. To post a waiting time, you need to know it. And waiting time is precisely the figure that counting at a door will never give you.
This article works through what a queue actually is, which four movements happen inside one, and why two of them leave no trace at a doorway. It names the five figures that make a queue manageable, and explains why measuring a queue is technically harder than counting an entry.
An entry is an event at a line. A queue is a state across an area.
Define it first, then measure it
This sounds like hair-splitting, and it is the reason so many queue figures are worthless: if there is no agreement on where a queue starts and ends, every system measures something different.
The literature is more precise than everyday language. A queue forms when more requests arrive per unit of time than the system can process - demand exceeds capacity. The server is whatever completes the transaction: a till, a counter, a kiosk, a person checking tickets. The area immediately in front of it, where the next person waits, is the buffer zone.
Three situations that look like a queue and are not:
- A group moving forward at a steady pace is not waiting. It is walking.
- People winding through a stanchion maze without ever stopping have not queued - there was no wait.
- If enough servers are free that nobody has to pause, there is no queue. Queuing starts with the first pause.
For measurement this means two things: the system has to tell standing from walking, and it has to separate the buffer zone from the queue behind it. Both are questions about an area, not about a doorway.
In practice this becomes a setup step you cannot skip: a queue needs two zones - the area where people wait, and the area where they are served. A waiting time only exists in the relationship between the two. Define only one zone and you get a head count inside a rectangle, which is not a queue but a gathering.
Four movements - and a door counter sees two of them
People in a queue do not only move forward. The technical terms are clumsy, but the behaviour is familiar to everyone:
- Balking - someone sees the queue and never joins it.
- Reneging, or abandoning - someone joins and then gives up because it is taking too long.
- Jockeying - someone switches queues because the other one looks faster. A variant: a group splits across several queues.
- Collusion - one person holds a place on behalf of several.
The first two are the expensive ones. Someone who never joins and someone who walks away does not buy, does not eat, does not book.
And neither leaves any trace at a door. A shopper who walks in, sees the checkout queue and leaves without buying is one entry and one exit - a perfectly ordinary visit in the data. It shows up later in the conversion rate, but without a cause attached: the number drops, and nobody knows why.
The two movements that cost revenue are exactly the two that trigger nothing at the door.
There is a second consequence for analysis: all four movements break any model. Anyone deriving a waiting time from arrival rates and service times is calculating around them. They have to be measured, not estimated.
The calculation that justifies the effort
The optimal queue is not the shortest one. It is the one where two costs balance: the cost of extra staff - including whatever that staff is not doing meanwhile - and the cost of business lost when customers walk away.
Both sides can be quantified. One sits in the payroll. The other sits nowhere at all, as long as nobody counts the people who give up.
For a sense of scale, a widely quoted estimate: in the United States alone, people spend roughly 37 billion hours a year in queues - a figure cited in the Harvard Business Review and reported by Bloomberg in 2015. The same source names wait-related frustration as the leading reason retailers lose customers. That is an estimate rather than a measurement, and it comes from a different market; as a rough answer to whether the topic is worth the effort, it is enough.
What matters is the second half of the sentence. To move towards the optimum, you have to know continuously how far away from it you are. A quarterly report does not tell you. Neither does one Tuesday morning.
Perceived and measured waiting time are two different things
It is rarely the duration itself that annoys people. It is the uncertainty. Someone who knows it will take twelve minutes waits differently from someone who has no idea whether it will be two or twenty.
That leads to something counter-intuitive: publishing the waiting time shortens the waiting time. Not the measured one, which stays as it is. But fewer people give up, and the average wait for everyone else falls too, because the queue distributes differently.
That is the theme park case from the opening. Sensor manufacturer Xovis documents it for Freizeitpark Plohn: after real, continuously updated waiting times were displayed, queue abandonment at the busiest attraction halved, from 13.5 to 6.5 percent. The same data made visible what nobody would otherwise have noticed - that dispatch intervals had stretched from two minutes to four.
The supermarket version runs the other way: a shopper who knows the checkout is currently six minutes rather than the usual fourteen keeps shopping for longer.
A posted waiting time does not make the wait shorter. It ends the guessing - and the guessing is the part that hurts.
The type of queue decides which figure matters
There is no such thing as the queue. There are several designs, and they fail in different places.
| Design | Where you meet it | Where it breaks |
|---|---|---|
| One queue, one server (FIFO) | Supermarket till, small counter | One slow transaction blocks everyone behind it. When a second server opens, the newest arrivals are served first - not those who have waited longest |
| One queue, several servers | Big-box retail, bank counters | Moves people fastest but looks longest - and long queues cause balking |
| Several queues, several servers | Checkout zone, check-in | Uneven load and jockeying; without per-lane measurement you only see the average |
| Priority-based | Business counters, ticket holders | The preferred group flows, the other one backs up - and is often not measured separately at all |
| Virtual | Deli counter tickets, time slots | Only works with live throughput data; otherwise the slots you hand out are guesses |
| In sequence (tandem) | Security, then passport control | The second queue is fed by the first. Measure only one and you will see the problem in the wrong place |
The practical conclusion: “how long is the queue” can only be answered once it is clear which queue is meant. In a tandem arrangement, total time through the process is the number that counts - not the length of any single section.
The five figures
No more are needed, and fewer will not do.
| Figure | What it tells you | Which decision it supports |
|---|---|---|
| Length | How many people are queuing right now | Immediate action: open a second till, move staff |
| Waiting time | How long it takes from the back of the queue to the server | The number for the display - and the one a commitment such as “under five minutes” is checked against |
| Service time | How long a single transaction at the counter takes | Whether a long queue is down to demand or to the counter - the question that decides between more staff and a different process |
| Throughput | How many transactions a server handles per hour | Staff planning, comparison across sites and shifts |
| Abandonment rate | How many join and then leave | The only figure that quantifies the loss - and therefore justifies the investment |
Abandonment is the one most often missing and the hardest to substitute. Without it, any business case about queues is an estimate.
Measured, estimated, not measured - three states, not one
Each of these five figures can be in one of three states, and a usable system says which:
- Measured - computed from a minimum number of completed observations within a defined window. Only this is a reading.
- Estimated - derived from the configuration, meaning the service capacity and average service time you entered. Useful for a forecast, not for proof.
- Not measured - there is no reading right now.
Not measured does not mean zero. A system that displays both the same way turns an outage into good news.
This is not a footnote for the manual, it is the question of whether the numbers can be trusted. A display that reports “0 minutes” when a sensor drops out sends people into a queue it cannot see. The honest output in that case is “no data” - ugly, and correct.
The same distinction belongs in every report. An abandonment rate that comes out of a model is not the same thing as a counted one. Both are useful; only one of them carries an investment decision.
Why a queue is harder to measure than an entry
An entry is an event: someone crosses a line, the system adds one. A queue is a state: at any moment, people occupy an area, and that area is not painted on the floor.
Two problems follow that appear in no brochure.
The boundary problem
At the back of every queue stands someone who may belong to it and may simply be standing nearby. A system that counts that person in one second and out the next produces a length that jumps - known in operations as flickering at the queue boundary. Camera-based installations fail here regularly, and the result is not a wrong number but an unstable one. A display that alternates between eight and fourteen minutes is worse than no display at all.
The perspective problem
A queue is dense. People occlude one another, and they do so most severely at exactly the moment the number matters most. A side-mounted camera sees the front of the queue and guesses at the back. Overhead sensing with depth information sees heads as separate maxima - even when they are touching.
We compared the four people-counting technologies in detail elsewhere: 3D sensor, camera, infrared and lidar or Wi-Fi. For queues, every point in that comparison gets sharper, because density is the normal case rather than the exception.
Active and passive management - and what each needs
There are two ways to influence a queue, and the best operations run both at once.
Active means changing capacity: open another till, add a security lane, move staff. That needs length and waiting time in real time - and throughput afterwards to know whether it worked.
Passive means changing behaviour without touching capacity: post how long it takes, and where it is currently faster. That needs the same waiting time, but reliable enough to put on a public sign - and a display that refreshes itself every few seconds rather than showing a number from earlier.
The most effective form is not the wait per queue but a direct pointer to whichever one is shortest right now. It does the arithmetic for the visitor and spreads the load without anyone intervening.
An airport terminal with three security checkpoints shows both in one picture. Deciding how many lanes to open is active. A display naming the wait at the other two checkpoints, plus the walking time to reach them, is passive - passengers then distribute themselves.
That this scale is real is shown by a documented installation: at Paris Gare du Nord, Eurostar departures run through French and British border control, security screening and ticket validation - according to the sensor manufacturer, thirteen queues within less than 2,000 square metres, covered by 69 sensors. Thirteen queues in that little space cannot be separated by counting at an entrance and an exit.
What belongs in the tender
Seven questions that separate a display board from an instrument of control:
- Is waiting time measured, or derived from arrival rates and service times? A derivation calculates around balking and abandonment.
- Is the abandonment rate reported? Without it the benefit cannot be quantified.
- How stable is the figure at the back of the queue? Ask to see one minute of raw values, not an average.
- Is each queue reported individually, or only the zone? With several servers, the average is the number that says nothing.
- Is there both real time and history? Real time runs the day, history plans the shift.
- How quickly does a value arrive? For a public display, one minute is the upper limit - anything older contradicts what the visitor can see.
- Does the system distinguish measured, estimated and not measured? And what does the display show when a sensor fails?
How we do it
The ANALYSIT Counting System measures queues across the area, not just across the doorway. Each queue is set up with its waiting zone and its service zone separately, and that is what yields length, waiting time, service time, throughput and abandonment rate. Throughput is the figure that is counted directly. Waiting time, service time and abandonment rate come out of observed waiting episodes - an episode starts when someone enters the waiting zone and ends either at the service zone or as an abandonment. Where too few episodes exist, what you get is a flagged estimate rather than a measurement. Every figure carries its state with it, measured or estimated, and a missing value is reported as missing rather than as zero.
Because every minute is kept as its own record, real time and 730 days of history sit side by side: the display gets its value, shift planning gets its trend. Thresholds such as “under five minutes” can be stored, so a commitment is not just an intention but a number you can look up at the end of the month.
Counting is anonymous. No images, no facial features and no device identifiers are created - not even in the queue, where people stand closest together.
If you want to know which of your queues can be measured at all, and where the sensors would have to go, talk to us. We look at the floor plan, the ceiling height and the process ahead of the queue - and tell you which of these figures are realistic in your case before we write a quote.