Event KPI definitions that stop the same number meaning two things

Two people in the same post-show review both said attendance and both meant something different, and it took forty minutes to work out that they were not disagreeing about anything.

The audience director's figure counted everyone who came through a door. The operations lead's figure counted everyone whose badge produced a scan at a turnstile. Between them sat about two and a half thousand people who came in through a side entrance during the morning rush when the readers were switched off to clear the queue.

Neither of them was wrong. Neither of them could say what their own number counted without going back to the person who built the extract.

What does an event KPI definition have to contain?

A usable metric definition names four things. The population, the filters, the time window, and the source system.

Attendance means unique badges with an entry scan, excluding exhibitor staff and contractors, across the three public show days, from the access control feed. Population: badges. Filter: exclude exhibitor staff and contractors. Window: the three public show days. Source: the access control feed.

Every one of those four is a place where two people can silently disagree, and every piece of attendee analytics downstream inherits the disagreement. Leave any of them out and the definition is a label. Most organisations' definitions documents are a column of labels with a sentence of explanation next to each, and the sentence almost always describes the intent of the metric instead of its construction.

The test is mechanical. Hand the definition to somebody who has never seen your data and ask whether they could write the query. If they have to ask a question, the answer to that question is a missing slot.

There is a fifth thing I now insist on, which is the unit of count. Attendance counted distinct on badge identifier, on email address, and on resolved person gives three different numbers from the same rows, and that choice deserves a line of its own. Which key you count distinct on works through what each one does to the total.

The rules already exist, in a standard nobody in events has read

Data definition is a solved problem in one narrow sense: somebody wrote down what makes a definition good, and the rules travel.

ISO/IEC 11179-4:2004, the part of the metadata registries standard covering the formulation of data definitions, sets out requirements and recommendations for constructing definitions for data and metadata. They are unglamorous and they catch most of what goes wrong in an event metrics dictionary. Three of the requirements do the work here. A definition shall be stated in the singular. It shall state what the concept is, so a definition built entirely out of exclusions fails. And it shall not embed the definition of another data element inside itself. Two of the recommendations carry just as much weight in a metrics dictionary: a definition should be concise, and it should be precise and unambiguous.

The rule against embedded definitions is the one event teams break constantly. A definition that reads "attendance is all qualified attendees who checked in" has hidden two other definitions inside it, qualified and checked in, and neither of them is written down anywhere. Pull them out, define them separately, and reference them by name.

The singular rule matters more than it looks, too. Write the definition of one attendance record, then say the metric is a count of those records. Defining the aggregate directly is how time windows go missing.

Three definitions of attendance on one edition

Here is the same show, the same weekend, and three defensible definitions.

Definition A counts unique badges with an entry scan, excluding exhibitor staff and contractors, across the three public show days, from the access control feed. That gives 24,300.

Definition B counts unique badges with any scan anywhere on site, entry points, session rooms and exhibitor devices, with the same exclusions and the same window, from the access control feed joined to the lead retrieval feed. That gives 26,800.

Definition C counts unique badges issued and collected, all registrant types, across four days including the pre-show conference day, from the registration system. That gives 31,400.

The gaps decompose cleanly. B minus A is 2,500 people who were scanned somewhere on site but never at an entry reader, which is the side-door population plus a VIP entrance that has no reader on it. C minus B is 4,600, made up of 2,900 exhibitor staff and contractors, 1,100 badges collected only on the conference day, and 600 badges printed and collected that never produced a scan of any kind.

Check the arithmetic: 2,900 plus 1,100 plus 600 is 4,600, and 26,800 plus 4,600 is 31,400. The lowest and highest figures differ by 7,100, which is 29.2 per cent of the lowest.

Three numbers, all computed correctly, all describing the same show. The spread is not a data quality problem. It is a definitions problem, and it will reappear every edition until somebody writes the four slots down. When two teams walk into a meeting holding two of those figures, settling the disagreement before the report ships is a different job with a method of its own.

Name the variants instead of arguing about which one is right

The instinct after seeing that spread is to pick a winner. I would rather name all three and publish one.

Give each variant a distinct name that cannot be confused: entry attendance, on-site attendance, badges issued. Then nominate one as the reported figure and say so in the definitions appendix. The other two keep existing, keep their names, and appear in the derivation section behind the headline.

Naming the variants does two useful things. Somebody who needs the larger number for a legitimate purpose, a venue capacity conversation for instance, has a named metric to ask for instead of quietly recomputing attendance with different filters. And the day the show's access control layout changes, you can see which variant moved and which did not.

CEIR does a version of this. The CEIR Index, published through IAEE, reports professional attendance as one of its four components, alongside net square feet of exhibit space sold, number of exhibiting companies, and gross revenue. CEIR's separate Cost to Attract Attendees research, reported in 2014, gives median organiser spending on a per verified attendee basis. Professional attendance and verified attendance are two named ideas doing two jobs, and having both named is what stops either of them absorbing the other. What the evidence behind a verified figure has to look like is a subject of its own.

How do you tell whether a definition is any good?

Definitions are cheap to write and expensive to get wrong, so it is worth having a way of checking them that does not depend on the author's judgement.

The first is the two-analyst test. Give the written definition to two people who did not write it, ask them each to produce the number without talking to each other, and compare. If they differ, the definition is incomplete and the difference tells you which slot is missing. This costs an hour per metric and it is the only test I have seen that reliably finds the assumptions the author did not know they were making.

The second is the boundary case list. For each definition, write down five records that sit near the edge and say what the definition does with each. A person who registered but collected their badge on the conference day and left. A contractor who also registered as a visitor. A badge scanned at 08:52 when doors open at 09:00. A registration cancelled on day two by somebody who had already attended day one. A duplicate created because a name was spelled two ways.

Those five cases take twenty minutes and they will change the definition about half the time. They also become the test fixtures for whoever rebuilds the pipeline in three years, which is the point at which most definitions quietly stop being implemented as written.

The source system slot is where definitions rot

Populations and filters get argued about in meetings, so they tend to be written down. The source system slot gets filled in once and then decays silently, and it is the slot that breaks a series.

Your access control feed came from one supplier in 2023 and a different one in 2025. The old feed emitted one row per badge per reader touch. The new one deduplicates within a rolling ten-minute window before it emits anything. Your definition still says "from the access control feed" and it is now describing two different things three editions apart.

So write the source slot as the system, the feed, and the version or supplier, and put a date on it. From the access control feed supplied by the venue's turnstile system, vendor A until the 2024 edition and vendor B from the 2025 edition. When the count moves 4 per cent between editions and nobody changed the definition, that line is where you look first.

The same applies to a registration platform migration, which is the single most common cause of a definitional break that nobody documented, because migrations are run as IT projects and the reporting definitions are not on the project plan.

Where this stops

Writing definitions properly stops two people meaning different things by the same word. It does nothing about a measurement that is wrong.

If your entry readers miss 8 per cent of the people walking past them, definition A is a precise, well-documented, four-slot definition of a number that is 8 per cent too low, and having it written down beautifully makes it more likely to be believed. Definitions are about shared meaning. Coverage is a separate problem with separate evidence.

The second limit is effort against value. A full definitions dictionary for a show is somewhere between forty and a hundred metrics, and writing all of them properly is a real piece of work that nobody has budget for. Getting that dictionary into a shape an auditor can follow, and keeping its versions straight so last year stays comparable, are two further jobs beyond this one. I would write the four slots for the five metrics that appear on the summary page and the ten that appear in the derivation sections, and leave the rest as labels until somebody argues about one. A dictionary that covers fifteen metrics and is accurate beats one that covers a hundred and was filled in by an intern.

Take the metric your organisation argues about most, write the four slots for it on one line, and send that line to the two people who most often disagree about the number. If they both accept the line, you have a definition. If either of them adds a condition, you have just found the filter nobody had written down.