Dwell time measurement on a show floor and what each method assumes

The same report gave two dwell figures for the same feature area on the same afternoon. Forty-seven minutes on one page, thirty-one on another. Both came from the same platform, both were correct, and the difference was that one counted the time between crossing a boundary and crossing back out, while the other counted the time a radio signal stayed above a threshold.

Dwell time measurement has three common implementations and they answer three different questions. Nobody gets into trouble for using the wrong one. They get into trouble for putting two of them in the same report without saying which is which.

What are the three methods actually measuring?

Portal dwell. A zone has a boundary with readers on it. Dwell is the elapsed time between an inbound crossing and the next outbound crossing. Everything inside that interval counts, including the twenty minutes the person spent sitting on a bench inside the zone with their back to everything.

Proximity dwell. A sensor sits at a point, usually a beacon at a stand or a reader under a carpet. Dwell accumulates while the badge's signal is strong enough to be inside a nominal radius. There is no boundary, only a threshold, and the threshold moves as the signal fluctuates.

Trajectory dwell. A positioning system estimates a location every second or two. Dwell is the time spent inside a polygon while moving below a speed threshold. This is the only one of the three that can distinguish standing at a stand from walking past it, and it is the one that needs the most from the underlying technology.

Zafari, Gkelias and Leung surveyed the underlying techniques in IEEE Communications Surveys and Tutorials in 2019, covering angle of arrival, time of flight and received signal strength across WiFi, RFID, ultra wideband and Bluetooth. The point that survives from that survey for an organiser is that every technology gives you a position estimate with an error distribution, and dwell is computed downstream of that error.

Portal dwell and the boundary you drew

Portal dwell is the most defensible of the three because the events behind it are discrete and countable. A crossing happened or it did not.

Its weakness is the boundary. A zone the size of a hall gives dwell figures dominated by the fact that people were in the hall. A zone the size of a feature area gives something more interesting and needs readers on every way in, including the corner where the carpet stops and nobody thought a person would walk.

Two failure modes matter. A missed exit read leaves the interval open until the badge appears somewhere else, which can turn 47 minutes into two hours and put an implausible value into a distribution of plausible ones. And a boundary with an unmonitored gap produces intervals that never close at all. The read rate arithmetic behind both of those is worked through in RFID badge tracking, and the practical defence is to publish the share of intervals that had both a start and an end.

Portal dwell also cannot see what happened inside. An attendee who entered a 400 square metre feature area, walked straight through it to the exit, and stood outside talking for half an hour before leaving through the same door, produces a large dwell figure and no engagement at all.

Proximity dwell and the radius that moves

Proximity dwell trades the boundary for a threshold, and the threshold is where the assumptions hide.

Faragher and Harle measured this directly in the IEEE Journal on Selected Areas in Communications in 2015, using 19 Bluetooth Low Energy beacons across a testbed of roughly 600 square metres. At a density of one beacon per 30 square metres they reported position errors under 2.6 metres 95 per cent of the time. Thinning the network to one beacon per 100 square metres widened that to under 4.8 metres, and a comparable WiFi system in the same space came in under 8.5 metres.

Put those numbers against a stand. A six by six metre stand covers 36 square metres. A 4.8 metre uncertainty circle covers 72.4 square metres, twice the stand. Even the dense case, at 2.6 metres, covers 21.2 square metres, which is 59 per cent of the stand and easily enough to include the aisle in front of it and part of the neighbour.

Then consider what the dense case costs. One beacon per 30 square metres across a 9,000 square metre hall is 300 beacons to deploy, power, map and collect. At one per 100 square metres it is 90 beacons and 4.8 metres of error. That trade is the actual decision, and it should be made against the size of the zones you intend to measure rather than against a vendor's accuracy claim.

Signal fading does the rest. A body between badge and beacon drops the signal below threshold for a few seconds at a time, and each drop ends an interval. Raw proximity dwell for a person standing still at a stand for half an hour arrives as a handful of fragments with gaps between them.

The same person, two answers

One attendee, one feature area, one afternoon.

The portals recorded an inbound crossing at 13:02 and an outbound crossing at 13:49, giving a portal dwell of 47 minutes.

The beacon in the same zone recorded that badge above threshold in six fragments totalling 31 minutes, with 16 minutes of accumulated dropout scattered between them. Reported raw, that is 31 minutes, and it is 34 per cent below the portal figure for the same person in the same place.

Now apply a merge rule, which every proximity implementation needs and few document. Open an interval after 30 seconds of continuous presence, and close it only after 90 seconds of continuous absence. The six fragments collapse into two intervals totalling 44 minutes, because five of the gaps were shorter than 90 seconds and one was not.

Three figures from one afternoon: 47, 44 and 31 minutes. The raw data never changed. What changed was a boundary definition and a gap parameter, and neither of them appears in any report I have seen produced by a platform of this kind.

That is the case for publishing the parameters next to the metric. A dwell figure with a method name, a zone definition and a merge rule attached can be compared across editions. A dwell figure on its own cannot be compared with anything, including itself next year.

Which method should you use for which question?

Match the method to the boundary you can actually defend.

For a hall, a conference zone or a fenced feature area, use portal dwell. The boundary is physical, the readers have somewhere to stand, and the resulting figure survives a challenge because the events behind it are discrete.

For stand level attention, proximity dwell is the only affordable option and it should be reported as a relative measure. Ranking forty stands by proximity dwell is reasonable. Telling an exhibitor that visitors spent 4.2 minutes at their stand is not, because the 4.2 depends on a radius, a threshold and a merge rule that the exhibitor cannot see and you probably have not tuned.

For anything involving where people stopped in open aisle space, you need trajectory dwell and therefore positioning accuracy tighter than the features you are trying to resolve. If the aisle is three metres wide and your 95 per cent error is 4.8 metres, the system cannot tell which side of the aisle somebody was on, and a heat map built from it is showing you the error distribution.

Read every dwell figure against the time of day it came from. Dwell in the first hour of a show is short because people are moving with purpose, and dwell after 15:00 is long because the hall has emptied and conversations run longer. Comparing two stands whose traffic peaks at different times compares two different behaviours, which is why the arrival curve belongs beside any dwell analysis. Whether somebody came back a second time is a separate and often more useful signal, covered in repeat visitor scans.

Where this stops

None of the three methods measures attention. They measure the position of a plastic card.

A badge in a bag under a chair generates dwell. A badge on a coat over an arm, held at hip height, generates a weaker signal and less dwell than the same person with the badge on their chest. Two people standing in the same spot for the same twenty minutes produce different numbers because of how they wore a lanyard, and no amount of parameter tuning corrects for it.

The second limit is that dwell distributions are heavily skewed and the mean is a poor summary. A zone with a median dwell of nine minutes and a mean of nineteen has a long tail of people who left their badge somewhere, and reporting the mean hands an exhibitor a number that describes nobody. Report the median and an upper percentile, and state the count of intervals the figure was built from.

The third is comparability. Change the beacon layout, the merge rule or the zone polygon between editions and the series breaks, silently, in a way that looks like a change in visitor behaviour. Version the configuration alongside the data and keep it with the rest of your attendee analytics documentation.

Start this week by asking whoever supplies your dwell figures three questions: what the zone boundary is, what closes an interval, and how many intervals were dropped as incomplete. If nobody can answer all three, the number is not ready to go in front of an exhibitor.