Decomposing year over year attendance variance into causes you can act on

The board pack says professional attendance was 25,900 against 26,800, down 3.4 per cent. Underneath it, a sentence about a softer market and a competing event.

Nobody in the room can do anything with that. Down 3.4 per cent is a fact about the past with no lever attached to it. The show director will ask what to fix, and the honest answer from a single percentage is that we do not know, because the same minus 900 is produced by a show that lost its best customers and by a show that had a brilliant acquisition year on top of a collapse in returns.

Those two situations need opposite responses and cost different amounts of money. Telling them apart takes one join and about two hours, and decomposing year over year attendance variance this way is the single highest-return thing in attendee analytics that most teams have never done.

Split the change into three populations, not one percentage

Every attendee at this edition is in exactly one of two groups: they were at the last edition, or they were not. Every attendee at the last edition is likewise either back or gone.

That gives three populations and one identity. Retained plus new equals this year. Retained plus lapsed equals last year. The change between the years is new minus lapsed, always, with no residual term.

Run it on the numbers above. Of the 26,800 who attended last year, 7,700 came back. Retention is 7,700 divided by 26,800, which is 28.7 per cent. The lapsed population is 26,800 minus 7,700, which is 19,100 people who were here and are not. The new population is 25,900 minus 7,700, which is 18,200.

Check the identity: 18,200 new minus 19,100 lapsed is minus 900, which is the headline. The decomposition is exact by construction, which is the property that makes it worth arguing from.

Now do the same thing for the previous pair of editions, because a single year's decomposition has nothing to compare against. Last year's 26,800 was 7,200 retained plus 19,600 new, on a base of 25,600 the year before that. Retention into last year was 7,200 divided by 25,600, or 28.1 per cent.

Put the two side by side and the story inverts. Retention went from 28.1 per cent to 28.7 per cent, and the retained count went from 7,200 to 7,700, up 500. Acquisition went from 19,600 to 18,200, down 1,400. Minus 1,400 plus 500 is minus 900.

Attendance fell and retention improved. The entire decline, and then some, came from bringing fewer new people through the door. That is a marketing and channel problem with a budget attached, and it is invisible in the 3.4 per cent.

Which channel and which geography lost the people?

The second cut takes the two moving populations and splits them by how people arrived and where they came from.

The new population of 18,200 breaks down as 5,200 from exhibitor guest codes, 4,600 from email to the house file, 3,100 from paid search, 2,400 from paid social, 1,900 from trade media and partner promotions, and 1,000 that cannot be attributed at all. Compare each line against last year's 19,600 and the shape appears immediately.

In this example the guest code line fell from 6,400 to 5,200. That single line is 1,200 of the 1,400 acquisition decline. Paid search and paid social were flat within a few hundred. Email was up slightly. Trade media was down 200.

Then geography. Splitting the acquisition change into domestic and international gives minus 1,750 domestic and plus 350 international, netting to the minus 1,400. So the show grew its international audience and lost domestic first-timers, and the domestic loss is concentrated in codes issued by exhibitors to their own customer lists.

That is now an actionable finding with a named owner. Somebody needs to know why exhibitors distributed or redeemed fewer guest codes, and the answer is usually one of three things: the code allocation changed, the redemption journey broke, or the exhibitor communications went out four weeks later than usual. All three are findable in an afternoon, and sizing an allotment so the codes actually get used is A18's subject.

Why does the retention number look so bad?

A 28.7 per cent retention rate reads like a disaster to anybody who has come from a subscription business, and the first reaction in the room is usually that the number must be wrong.

It is not wrong. Freeman's End-of-Year Trends Recap on 2025, released in January 2026, puts the industry's blended attendee retention rate at barely above 30 per cent, which the recap translates as 70 per cent of people not coming back. It adds that the best and most generous data it could find, looking at trends since 2008 in three-year rotations, puts the figure in the low forties. So a show sitting at 28.7 per cent is somewhat below a weak industry norm rather than uniquely broken.

Two consequences follow, and they are the reason this decomposition is worth building rather than admiring.

The first is that acquisition at a trade show is the maintenance activity, whatever the budget line calls it. At 30 per cent retention you have to recruit roughly seven tenths of your audience every year to stand still. A budget conversation that treats attendee marketing as discretionary is treating replacement of the audience as discretionary.

The second is that a point of retention is worth more than a point of acquisition, because retained attendees cost nothing to find and, on most shows, convert better for exhibitors. Moving retention from 28.7 to 31.0 per cent on a base of 26,800 is 616 more people, and it is the cheapest 616 people you will ever add. Whether that is achievable is a different question, and the shape of the new versus returning attendee mix by year of first attendance, which is where the room to move actually is, is D15's argument.

Two things this decomposition will get wrong if you let it

The join that produces the retained population is an identity match, and identity matching on a registration file is imperfect in a direction that flatters nobody.

Match too loosely and you merge two different people, which inflates the retained count and deflates both new and lapsed. Match too tightly, which is the more common failure, and you count a returning attendee who changed employer as a lapsed person plus a new person. That double error moves both of the numbers you are trying to read, and it moves them in the direction that makes your acquisition look better and your retention look worse than they are.

The practical control is to run the decomposition twice, once on exact lowercase email and once on your fuller matching rules, and report the gap. If retention is 24.1 per cent on email alone and 28.7 per cent on the fuller rules, the 4.6 point difference is the part of your answer that depends on the matcher, and it should be stated rather than absorbed.

The second thing to guard against is reading any of this as performance before correcting for the calendar. A show that moved two weeks later, ran three days instead of four, or lost a day into a school holiday will produce an acquisition and retention split that looks like a marketing story and is really a date story. Isolating the calendar shift effect on attendance is D16's subject and it belongs upstream of everything in this post. The same decomposition run on registrations before doors open, where the pacing window itself moves, is A5's year over year registration comparison.

Publish it as four numbers and a bridge

The output that survives a board meeting is small.

Last year 26,800. Retained 7,700. Lapsed 19,100. New 18,200. This year 25,900. Underneath, the two prior year comparators: retention 28.1 per cent then 28.7 per cent, acquisition 19,600 then 18,200.

That is nine numbers, they reconcile exactly, and anybody can check the arithmetic in their head. Put the channel split of the acquisition change on the facing page and the meeting is about the guest code programme within about ninety seconds, which is where you wanted it.

Compare that with the alternative, where somebody spends the meeting defending 3.4 per cent against a market backdrop, which is a discussion that has never once produced a decision. The CEIR Index, released by IAEE on 4 May 2026, forecasts 2.1 per cent growth for the total index in 2026, and knowing that is genuinely useful context for expectations. It is useless as an explanation of your own 900 people.

Where this stops

The decomposition is arithmetic, and arithmetic does not know why anybody did anything.

It will tell you that 19,100 people did not come back. It will not tell you whether they retired, changed industry, sent a colleague instead, went to a competitor, or simply had a diary clash. The colleague case is the one that quietly distorts everything, because at a company level the relationship is intact and at a person level it reads as a lapse plus an acquisition. If your show sells to companies rather than individuals, running the same three-way split with company as the key, alongside the person-level version, is worth the extra hour and will usually produce a materially calmer picture. On the exhibitor side of the file the equivalent account-level view is footprint change year over year, which is F31's.

The other limit is that lapsed is not a decision you observed. A trade show is not a subscription, nobody cancels, and somebody who skips one edition and returns for the next was never really gone. On a show with a two year buying cycle in part of its audience, a meaningful share of your 19,100 will be back next time, and treating them as churn will send an expensive win-back campaign at people who were always coming.

The honest version of this is to publish the consecutive-edition split as the headline, because it is simple and reproducible, and to add one line underneath giving the share of the previous edition's attendees who appeared at either of the next two. On most shows that second figure runs several points higher and it is the one to use when sizing a win-back budget.

Take your last two editions, join on your best identity key, and produce the three counts. If retained plus new does not equal this year's attendance exactly, your join is wrong and nothing downstream of it is worth reading.