The unique attendee definition decides whether your show grew or shrank

An audience director walked into a review with growth of 4.1 per cent and walked out with a decline of 0.5 per cent, and nothing about the show had changed in the interval.

What changed was the key. Her figure counted distinct badge records. The analyst who had rebuilt the pipeline over the summer counted distinct resolved people. Same registration file, same edition, same three days, and the two numbers moved in opposite directions against the prior year.

Nobody had made an error. The show had genuinely grown on one key and genuinely shrunk on another, and until somebody wrote down which key the headline used, both statements were true. The unit of count is the fifth slot in a metric definition, and it is the slot most often left blank.

What can a unique attendee count be unique on?

Unique on what is the whole question, and there are four answers in normal use. Counting unique attendees against visits is a separate axis, and it multiplies with this one.

The badge record. One row per badge issued, which is what most registration platforms hand you and what most reports quietly use. A person who registers twice, or who loses a badge and gets a reprint, appears twice.

The email address. Lowercased, trimmed, sometimes with the plus-suffix stripped. It collapses the double registrations and it fails on anyone with more than one address, which is disproportionately the senior buyers who changed jobs between editions.

The resolved person. A record produced by matching across name, email, phone, company and whatever else you hold, with a probabilistic score and a merge decision behind it. It is the closest thing to counting human beings and it is the only one of the four that requires a method rather than a column. That method is the identity resolution cluster's subject, and the thing to carry into a reporting conversation is that the resolved person count depends on where the merge threshold was set.

The organisation. One row per resolved company. Useful, entirely different in scale, and it belongs on its own line rather than in a sentence about attendance.

Publish the whole ladder every edition

The move I would make is to compute all of them and publish the ladder, in the derivation section behind the headline, every edition.

Four rungs, four counts, and a one-line explanation of each gap. It costs one extra query. It ends the argument permanently, because anybody who wants a different key can read their number off the ladder instead of quietly rebuilding the extract with different logic and circulating a fifth figure. Where the argument has already started, a definition diff is the faster route out of it.

It also makes the pipeline's behaviour visible. When the gap between badges and emails doubles in one edition, something changed in how registrations are being submitted, and that is worth knowing in November rather than the following August.

The ladder on two editions

This edition first. The registration and access control data yields 33,200 badge records that met the entry criteria. What counts as meeting them depends on the evidence behind each row, which is a decision separate from the key.

Keying on email drops that to 30,400. The 2,800 difference is 1,900 badges that were the second or later badge for an email already counted, mostly group registrations re-entered by an assistant plus a run of reprints on day one, and 900 badges that carry no email address at all and get dropped by the email-keyed query.

Keying on resolved person drops it again to 27,800. That 2,600 is people holding more than one email address: 2,400 people with two and 100 with three, which is 2,400 plus 200, giving 2,600.

Keying on organisation gives 15,100. Divide 27,800 by 15,100 and the show averaged 1.84 attendees per company.

The top and bottom rungs differ by 5,400, which on 33,200 is 16.3 per cent.

Now the prior edition, computed the same way: 31,900 badges, 29,800 emails, 27,950 people.

Badge-keyed growth is 1,300 on 31,900, up 4.1 per cent. Email-keyed growth is 600 on 29,800, up 2.0 per cent. Person-keyed growth is minus 150 on 27,950, down 0.5 per cent.

The cause sits in the first gap. Duplicate badges per email went from about 1,100 to 1,900 between editions, because a group registration flow was promoted heavily and assistants entered their delegates one at a time without the system recognising repeats. The show sold more badges to slightly fewer people, and only the person-keyed count can see that.

Whether it matters depends on what the number is for. For a venue capacity conversation the badge count is the right one. For the board's audience trend, the badge count is telling them the opposite of what happened.

Changing the key changes about nine other numbers

The reason this decision is worth a meeting is that the attendance headline is a denominator almost everywhere else in your attendee analytics, so moving the key moves a column of figures that nobody thought they were changing.

Work through the same edition. Cost per attendee divides acquisition spend by the count, so switching from badges to people raises it by 19.4 per cent, because 33,200 divided by 27,800 is 1.194. The exact treatment of that metric is D19's, and the point here is only that the denominator swap happens silently. Attendance per exhibiting company falls by the same factor. The registration to attendance conversion rate falls, because the numerator shrinks while the registration base usually does not. International share moves, sometimes by a full point, since duplicate badges are not evenly distributed across markets.

The new versus returning split, which D15 covers, does not merely move. It becomes computable. On a badge key you cannot distinguish a returning attendee from a person who registered twice this edition, so the returning share is not a measurement at all.

So change the key once, restate the prior edition at the same time, and reissue the affected figures together. Changing it quietly, in one report section, while the rest of the pack still divides by badges, produces a document whose internal ratios no longer multiply out. Somebody will notice, usually an exhibitor with a spreadsheet, and the question they ask will be about the whole report rather than about one line.

Approximate counts are a separate trap

If your reporting layer computes distinct counts over hundreds of millions of rows, there is a reasonable chance it is not counting them exactly.

Flajolet, Fusy, Gandouet and Meunier published HyperLogLog at the 2007 conference on analysis of algorithms, and it is now the standard way large systems estimate cardinality. Their result is that with an auxiliary memory of m units the relative standard error is about 1.04 divided by the square root of m, and that cardinalities beyond a billion can be estimated to a typical accuracy of 2 per cent using a kilobyte and a half.

Two per cent is superb engineering and it is larger than most of the year-on-year movements an organiser reports. A 0.5 per cent decline computed by an approximate distinct count is not a measurement of anything.

Event data is small by the standards that algorithm was built for. A portfolio with eight shows and five years of history has a few million registration rows, and an exact count-distinct on that is trivial. So check whether the distinct counts on your reporting surface are exact, and if the answer is that the warehouse or the BI tool is doing something clever for speed, turn it off for the published figures. Approximation belongs in exploratory work.

Which key should the headline use?

I would report the resolved person count as the headline and carry the badge count on the ladder beneath it.

The argument is about what the number is used for. Every downstream question an organiser asks is a question about people. Did this year's audience come back. What share were first-timers. What did it cost to bring one of them. All of those need a person key, and computing them off badge records gives answers that are wrong in a consistent direction, because the duplicate badges are concentrated in group-registered corporate delegations who behave differently from individual registrants.

The Freeman Trends Report published in 2025 reports that only about 30 per cent of first-time attendees return to an event. That statistic is only computable against a person key. On a badge key you cannot tell a returner from a person who registered twice this year, and a portfolio-wide retention figure built on badges will overstate new attendance every edition.

The counter-argument is real. Resolved person counts depend on a matching threshold that somebody chose, and moving the threshold moves the headline, which makes the metric feel less solid than a count of rows. My answer is that the badge count also depends on choices, in the registration flow, made by people who were optimising for conversion and were not thinking about your series at all. At least the matching threshold is written down.

The CEIR Index, published through IAEE, names professional attendance as one of its four components. It does not tell you which key to count on, and it is not trying to. That decision is yours, and the only failure is leaving it unmade.

Where this stops

The resolved person count inherits every weakness of the matching that produced it, and those weaknesses are not symmetric. Two colleagues at a small company sharing a surname and an address will merge, and no threshold fixes that. A person with sparse fields will never merge with anything, so registration channels that collect fewer fields will look like they bring more unique people than they do.

That second effect is worth sizing before you use the person count for channel comparison. If your self-service form collects four fields and your telesales route collects nine, the self-service route will appear to deliver a higher share of unique attendees, and the difference is a measurement artefact. Nothing has been shown about how the two routes really perform.

The organisation rung has its own problem. Company name text is a poor key, and the gap between 15,100 organisations and the truth is usually large. I would publish that rung with an explicit note that it is name-based unless you have run proper company resolution.

Run the ladder on your last edition. Four queries, four counts, one afternoon. If the badge-to-person gap is above about 10 per cent, whichever key your headline currently uses is making a decision for you that nobody has approved.