Survey length and completion rate decide how much show feedback is real

Three thousand one hundred and forty people opened the post show survey. One thousand three hundred and nineteen reached the end. The report will quote the finishers and describe them as attendee feedback.

Survey length and completion rate get discussed as if the relationship were a smooth curve you slide along. It behaves more like a staircase. Most of the loss happens at two or three specific items, and once you know which ones, the decision about what to cut becomes concrete rather than a negotiation about how many questions feels reasonable.

Where the completions actually go

Run the survivor count per item. Most survey platforms will export it and if yours will not, the response table has a timestamp per item and you can count non-nulls per column.

An eighteen item instrument on a mid-sized show produced this shape.

Completion is 1,319 over 3,140, or 42.0 per cent. Total loss is 1,821 people.

Three items account for 478 plus 275 plus 368, which is 1,121, or 61.6 per cent of everything lost. The remaining fifteen transitions cost 700 between them, an average of 47 each. Anyone arguing about whether the survey should have fifteen or eighteen questions is arguing about the 47s.

The grid at item six is the expensive one, and it is expensive twice over. It costs 478 people directly, and because it sits at position six it costs those 478 people their answers to everything from item seven onwards. Moving the same grid to position fifteen would still lose respondents, and it would lose them after they had answered the twelve items you care more about.

Is a shorter survey always better?

The industry answer is yes and the aggregate data says something more complicated, which is worth taking seriously before you cut on instinct.

Survicate analysed 21,863 surveys collected across more than fifty industries through 2025 and published the completion curve by item count in March 2026. A single question completed at 85.7 per cent, two at 80.7, three at 77.4, five at 72.5 and six at 69.4. So far the story is the expected one. Then it stops behaving. Twenty questions came in at 68.7 per cent, the 21 to 40 band was the worst point on the whole curve at 70.5 per cent, and instruments of 101 to 150 questions completed at 83.0 per cent, better than a four question survey.

The response side moves the opposite way. Survicate recorded 33.9 per cent response for one to five item surveys against 57.7 per cent for the 101 to 150 band, a pattern it attributes to casual clicking on short instruments.

The reason the long end recovers is selection. Nobody finishes a 120 item questionnaire by accident. Those respondents are panel members who were told the length, accepted it and were probably paid. A post show survey sent to an attendee list has none of that filtering, so your show sits in the part of the curve between five and forty items where the aggregate data is at its worst, and every item you add is taken out of completions.

I would draw one practical conclusion from the shape of that curve. There is no length at which a post show survey becomes long enough to earn the selection effect back. The choice is between a short instrument and a middling one, and middling is measurably the worst place to be.

What a long questionnaire does to the answers that survive

Completion is the visible cost. The invisible one is what happens to the answers people do give.

Galesic and Bosnjak reported in Public Opinion Quarterly in 2009 on an experiment that manipulated both the stated length of a web questionnaire, at 10, 20 and 30 minutes, and the position of blocks of questions within it. The longer the stated length, the fewer respondents started and the fewer completed. Questions positioned later in the questionnaire drew faster responses, higher item non response, shorter answers to open ended questions, and less variation across items arranged in grids.

That last finding is the one that should change how you read your own data. Straight-lining down a grid at item fifteen looks identical in the export to a considered set of consistent ratings. If your operations scores all cluster at four out of five and they sit late in the instrument, you may be reading fatigue as consensus.

The consequence for a show report is direct. An eighteen item survey does not give you eighteen items of feedback from 1,319 people. It gives you maybe eight good items and ten that degrade as the instrument runs on, from a group who were willing to sit through all of it.

Setting a usefulness threshold per item

The cut is easier once each item carries a price.

Take the survivor table and assign each item the completions it costs. The grid costs 478. The first open text costs 275. The second open text costs 368. A single choice item costs somewhere around 47.

Then ask, item by item, one question with a checkable answer: did any figure derived from this item appear in the last two post show packs, or drive an operational change? Where the answer is no, the item costs completions and returns nothing, and it goes. Where the answer is yes, weigh the price. An item costing 47 people needs only to be occasionally useful. An item costing 478 needs to be something a named person acts on every edition.

Applied to the instrument above, the satisfaction grid failed the test. Its fourteen rows produced one chart nobody had referenced in two years, and it was replaced by two single choice items covering the two rows that had ever been cited. The second open text asked what the attendee would change about the conference programme, which the programme team had never read, and it went entirely.

The test has a useful side effect. Running it forces somebody to name an owner for each surviving item, and an item with no owner is an item that will produce a chart with no reader. I would run the exercise with the operations lead and the programme lead in the room rather than by email, because the honest answer to "did you use this last year" is easier to give out loud than to write down.

Two objections come up every time. The first is that an item has to run for a few editions before its value shows, which is fair for a genuinely new measure and is a reason to give it a fixed trial of two editions with a decision date, rather than permanent residency. The second is that cutting an item breaks the year-on-year series. It does, and the series it breaks is one nobody was reading, which is why the item failed the test. Where a series is genuinely load bearing, keep the item and pay its price knowingly.

What does cutting eleven items actually buy?

Suppose the next edition runs seven items on 3,100 starts and completes at 78 per cent. That is 2,418 finishers against 1,319, an increase of 1,099 usable responses on a slightly smaller start.

The arithmetic that matters is what those 1,099 do to the rest of your reporting. More finishers means a smaller non response rate, which shrinks the first term in the bias identity and makes the mean more defensible, and sizing that effect is post show survey response bias in D23. It also means the demographic cells inside your attendee analytics breakdowns have enough responses to report separately, which is usually where a small survey fails first. A cell with 40 responses cannot carry a satisfaction figure. The same cell with 73 can.

Deciding which seven items survive is a design question rather than a length question, and post event survey design in D22 argues that most of the room comes from deleting whatever the registration record already holds.

Where this stops

Cutting items raises completion. It does nothing about who chose to open the invitation in the first place, and that is the larger of the two selection effects at most shows.

The instrument above lost 58 per cent of the people who started. It never saw the twenty thousand who did not start at all, and no amount of shortening reaches them. If your open rate is eleven per cent and your completion is forty-two, the shortening exercise is competing for the smaller half of the problem. When you press send is the other lever, and that sits with post event survey send timing in D25.

There is also a floor below which cutting becomes counterproductive, and Survicate's response figures point at it. A two item survey collects clicks from people who have nothing to say, and a mean built on casual clicking is worse than a smaller mean built on considered answers. Somewhere around five to eight items you are trading real feedback for volume, and the honest answer about where that line sits for your audience is that you will have to test it.

Export the response table from your last post show survey and count the non-nulls in each column. Put those counts in a single column of a spreadsheet, look at the two largest drops, and find out what those two items produced in the last report. That is a twenty minute exercise and it will name the questions to delete before you write the next instrument.