Verified attendance reporting and the evidence trail behind every headline number

An operations lead was asked, in a meeting, where the attendance figure came from. He said the badge system. He was asked which part of the badge system and he could not answer, and neither could the analyst who had built the extract, because the extract read a view that somebody had created two years earlier and nobody had opened since.

The number was probably fine. That was the problem. Probably fine is the state most attendance figures are in, and it holds up right until somebody with a reason to doubt it asks a second question.

The fix is not a better number. It is a column.

What does verified attendance actually have to mean?

Inside your own building, verified should mean one thing: every row counted in the attendance figure has a physical event behind it that was recorded at the time it happened.

A badge read at an entry reader. A badge printed and collected at a kiosk. A turnstile count matched to a badge. Those are events with timestamps. A row in the registration table saying somebody intended to come is not one, and neither is a row created afterwards by a person typing into a spreadsheet, though that second category can still be counted if you say so.

The word already has currency in the sector. CEIR has used verified attendance as a reporting basis for over a decade: its Cost to Attract Attendees research, reported in Exhibit City News in 2014, gives median organiser spending on a per verified attendee basis and breaks results out by exhibition metrics including gross revenue, verified attendance and net square footage of paid space. Separately, the CEIR Index published through IAEE tracks professional attendance as one of its four components. Two named metrics, doing two jobs, and the existence of both is the reason your own figure needs to say which idea it is implementing.

What a third party would require before signing anything is a different subject with its own rules, and what an attendance audit actually checks and the gap between verified and registered attendance handle it. Everything here is about what you owe yourself. It sits on top of the key you count distinct on, which decides what a row is before evidence decides how far to trust it.

Tag every row, from a closed list

The mechanism is one column on the attendance fact table: evidence type, populated for every row, drawn from a short closed list.

Four values covers most shows. Scanned, meaning a badge read at a reader with a timestamp and a device identifier. Printed, meaning a badge produced and collected at a named kiosk, with no subsequent read anywhere. Matched, meaning a physical count reconciled to a badge population by a rule you can state. Manual, meaning a human asserted the row.

The list has to be closed and short. A free-text provenance field decays into forty variants within two editions and cannot be reported on. If a new evidence type genuinely appears, adding a fifth value is a change to the report specification with a changelog entry, which is D2's mechanism.

There is an established way to think about this. The W3C published PROV-DM, the PROV data model, as a Recommendation on 30 April 2013, and it defines provenance as a record describing the people, institutions, entities and activities involved in producing a piece of data, usable to form assessments about its quality and trustworthiness. Its core concepts are entity, activity, agent, derivation and attribution. An evidence-tagged attendance row is a small instance of exactly that: the entity is the counted attendance, the activity is the scan, the agent is the device or the person, and attribution is the column you just added.

You do not need the full model. You need its instinct, which is that the record of how a value came to exist is part of the value.

The mix, on one edition

A show reports attendance of 19,400. The evidence column gives 16,300 scanned, 2,130 printed and never scanned, and 970 manual, with no rows in the matched category because this venue has no turnstile counts.

As percentages: 16,300 divided by 19,400 is 84.0 per cent, 2,130 is 11.0 per cent, and 970 is 5.0 per cent. The three sum to 19,400 and the percentages sum to 100.

Read what that mix says. Eleven per cent of the headline is people who collected a badge and then never passed a reader, which is either a coverage gap at the entrances or a real population who took a badge and left. Whether that population belongs inside the attendance figure at all is a ledger question, and the registration to attendance reconciliation is where it gets decided line by line. Five per cent is people whose attendance exists because somebody said so.

Neither of those makes the number wrong. Together they mean 16 per cent of the figure rests on something other than a device reading a badge, and the person quoting 19,400 in a sales meeting should know that before an exhibitor asks.

The mix is also the diagnostic that improves year on year. If your scanned share moves from 84 to 91 per cent because you put readers on the two side entrances, that is a real improvement in the measurement, and it will look like an attendance increase unless you report the mix alongside the total.

Publish the mix next to the number, in the same block

One line under the headline. Attendance 19,400, of which 84.0 per cent scanned, 11.0 per cent badge collected without a subsequent read, and 5.0 per cent manually recorded.

That sentence costs eleven words of report space and it does two things. A reader can calibrate how much weight to put on the figure without asking anybody. And the people producing the number acquire a visible incentive to improve the mix, which is the fastest route I know to getting a reader installed on the door that has never had one.

The objection is that disclosing 5 per cent manual makes the show look weak. I disagree, and the reason is what happens when you do not disclose. A figure presented as a clean count gets challenged as a whole the moment anybody finds one soft row, and then you are defending 19,400 rather than defending 970. Publishing the mix converts a possible credibility failure into a stated limitation, and stated limitations are dull, which is what you want. It is also what lets a reader tell an organiser reported figure from an audited one without having to ask.

Populating the column is an operations job

The evidence tag has to be written when the row is created, by whatever process created it, and this is where most attempts at this fail.

An analyst reconstructing provenance three weeks later is guessing. They can see that a row has no scan and infer that a badge was printed, but they cannot distinguish a badge printed and collected from a badge printed and abandoned on the kiosk shelf, and they certainly cannot tell you which of forty manual rows came from a group check-in and which came from somebody correcting an error.

So the change belongs upstream. The registration platform stamps its own rows. The access control feed stamps its own. The on-site terminal stamps its own. Manual entries require a reason code chosen from a list of five, entered by the person making the entry, at the time. That last requirement is the only one that meets any resistance, and it takes about four seconds per row.

Name an owner for the column, which should be operations rather than the reporting team, because operations controls every system that produces the underlying event. A reporting team that owns the tag ends up inferring it, and an inferred provenance column is worse than none, since it looks like evidence. That split of ownership is unusual in attendee analytics, where the reporting team normally owns every column it publishes.

What happens when the number changes after publication?

Attendance figures move after publication. Terminals sync late, duplicate rows get found, a manual list turns out to have been entered twice. Where the movement is allowed to stop is a freeze date question. What happens to figures already in circulation is this one.

Three weeks after the show above published 19,400, two things happened. A registration terminal that had failed to sync uploaded its scan log, converting 210 rows from printed to scanned. And a reconciliation found that 640 of the 970 manual rows were duplicates of people already counted as scanned.

The restated figures are 16,510 scanned, 1,920 printed only, and 330 manual, totalling 18,760. The headline falls by 640, which on 19,400 is 3.3 per cent. The mix improves to 88.0 per cent scanned, 10.2 per cent printed only, and 1.8 per cent manual.

A revision policy decides in advance what happens next, and it needs three things. A materiality threshold, expressed as both a percentage and an absolute count, since 1 per cent of a small show is not worth a republication and 1 per cent of a large one is thousands of people. A named person who approves a restatement. And a rule that the superseded figure stays visible, with the date it was published and the date it was superseded.

I would set the threshold at the larger of 1 per cent and 250 attendees. The 640-person revision above clears it, so the report is reissued with a revision note, the prior version stays in the archive rather than being overwritten, and anybody who quoted 19,400 in the intervening three weeks can find out why their number is now different.

The alternative, which is what most organisations do, is to silently correct the figure in the next version of the file. That is how two attendance numbers for the same edition end up in circulation with nothing to say which is current.

Where this stops

Evidence tagging tells you what kind of record sits behind each counted person. It cannot tell you whether the record is right.

A reader that double-counts a re-entry produces a scanned row with a timestamp and a device identifier, which is the highest-confidence tag in your list, for a person who was already counted. Provenance and accuracy are different properties, and a well-tagged wrong number is still wrong. Reconciling counts across systems that each claim a different total is D11's subject, and it is the check that sits next to this one.

The second limit is that the manual category will never go to zero, and chasing it there is a poor use of a season. Group check-ins for a delegation of forty, a VIP entrance run by a person with a clipboard, a session room where the reader was moved: these produce manual rows for good operational reasons. The target is not elimination. The target is that the manual share is small, known, and stable enough that a movement in it means something.

Add the evidence column to your attendance table for the last edition, populate it from whatever you can reconstruct, and count the rows you cannot classify at all. That count, the rows with no determinable provenance, is the honest measure of how much of your headline you can currently defend.