A brand lift study for sponsors needs a control group you can actually build

The sponsor asks whether people remembered them. You run a post-show survey, 400 people answer, and 62 per cent of them pick the sponsor's name from a list of companies associated with the event. That goes in the report as a headline.

Three days later the sponsor's agency sends one line back. Compared with what?

It is the right question, and a brand lift study for sponsors is the design that answers it. The word lift only means anything against a baseline, and a single percentage from a single group of respondents has no baseline in it anywhere.

What is the comparison you are missing?

Sixty-two per cent recall could mean the signage worked. It could also mean the sponsor is the largest firm in the sector and 62 per cent of your audience would have named them at an event they never attended.

The counterfactual is the whole measurement. You need two groups answering the same question at the same time, one that met the sponsorship and one that did not, and the difference between them is the only part attributable to what the sponsor bought.

Consumer brand lift studies buy this with random assignment: a holdout group is deliberately excluded from the campaign. On a show floor you cannot do that. You cannot randomise who walks past a hall banner, and you would not want to sell a sponsor an asset that a randomly chosen third of the audience is prevented from seeing.

So the comparison group has to come out of the exposure data you already collect, which makes this an observational study with all the honesty problems that implies.

Splitting the audience with data you already hold

Three splits are usually available without adding anything to the survey.

Badge scans by zone. If the sponsorship is hall-specific, the set of people whose badge was read by a scanner in that hall is your exposed group and the people who only ever scanned elsewhere are the comparison. This works cleanly for a show with distinct halls and poorly for one open floor.

Session attendance. If the sponsor took a keynote or a theatre block, attendance at that session is a hard exposure flag, and it is usually the strongest one available because the exposure was sustained and in front of a seated audience.

App events. A sponsored card, an interstitial or a directory placement produces a server-side record of who the thing was rendered to. That gives an exposure flag at the individual level with a timestamp, which none of the physical assets can match.

Whichever split you use, write down what else separates the two groups. People who attended the sponsored keynote are more interested in that topic than people who did not, and interest in the topic predicts recall of a company in that space regardless of any banner. That confound does not disappear because you ignored it.

The arithmetic on a 400 response survey

Suppose the survey returns 400 usable responses, 200 of whom scanned into the hall carrying the sponsor's branding and 200 of whom did not. Aided recall of the sponsor comes back at 62 per cent in the exposed group and 41 per cent in the unexposed group.

The point estimate of the lift is 21 percentage points. Now the interval, which is what stops the number being oversold.

The standard error of a difference between two proportions is the square root of the sum of each group's variance term. For the exposed group, 0.62 times 0.38 divided by 200 is 0.001178. For the unexposed group, 0.41 times 0.59 divided by 200 is 0.0012095. Add them to get 0.0023875, and the square root is 0.0489, or 4.9 percentage points.

Multiply by 1.96 for a 95 per cent interval and the margin is 9.6 points. So the honest statement is a lift of 21 points, with a 95 per cent interval running from 11.4 to 30.6.

That is a usable result. The interval excludes zero comfortably, so the difference is real at conventional confidence, and the width tells the sponsor exactly how much precision they bought. A report that prints 21 and stops has thrown away the second half of the finding.

How small a lift can a show survey detect?

Turn the arithmetic around and it becomes a planning tool.

Hold the two proportions near 0.62 and 0.41 and ask what sample gives a margin of 5 points instead of 9.6. The variance sum for one respondent in each group is 0.2356 plus 0.2419, which is 0.4775. Set 1.96 times the square root of 0.4775 divided by n equal to 0.05, and n comes out at 734 per group, so 1,468 completes in total.

Now put that against a real show, using whatever response rate your own post-show survey achieved last edition. Take 8 per cent as the working assumption and substitute your figure. At 8 per cent, 1,468 completes needs 18,350 invitations, which is more people than most shows have in the building. Run the same calculation at a 6,400 attendee show and the arithmetic goes the other way: 8 per cent of 6,400 is 512 completes, roughly 256 per group, and the margin at that size is 8.5 points.

The conclusion is worth stating plainly because it saves people a great deal of wasted effort. A show-scale brand lift study can establish that a twenty point difference exists. It cannot resolve a five point one, and no amount of careful wording in the report changes that. If the sponsor's objective is a small movement in preference, the study they need is a market-wide brand tracker with a different budget, and the answer is to say so at contract stage.

Why the recall question itself moves the number

The way you ask is part of the measurement.

Unaided recall asks the respondent to name the sponsors with no prompt and returns low figures. Aided recall shows a list and asks which of these sponsored the event, and returns much higher ones. Both are legitimate, and mixing them across editions makes a trend line meaningless.

Wakefield, Becker-Olsen and Cornwell examined the effects of sponsorship level, prominence, relatedness and cueing on recall accuracy in the Journal of Advertising in 2007, which is a useful reminder that the question format and the fit between sponsor and property are variables in their own right. Change the list length, the order or the category framing between years and you have changed the instrument, so the comparison you publish is partly about your questionnaire.

One cheap control fixes the most common failure. Put a plausible non-sponsor on the aided list, a company in the same category that bought nothing. Some respondents will select it, because they associate the category with the event and are guessing. If 12 per cent of the unexposed group picks the decoy, then 12 points of your 41 per cent baseline is noise of exactly that kind, and the sponsor deserves to see it. Report the decoy rate every year and you also get a stability check on your own sample.

What the exposure flag is actually measuring

Breuer and Rumpf used eye tracking to study how viewers process sponsorship information in sport telecasts, in the Journal of Sport Management in 2012, and found that the capture of attention is determined by the placement of sponsor signage and by exposure variables, with sponsor recall in turn a function of attention and brand-related variables.

Carry that across to a show floor and the binary exposed flag looks crude. A person who walked under the banner once on the way to a session and a person who queued beneath it for twenty minutes both count as exposed, and the second person had a different experience of the same asset. Recall is downstream of attention, and attention is downstream of placement and duration.

The fix is not complicated. Where the scan data supports it, band the exposed group by intensity: one scan in the zone, two to four, five or more. Then report recall by band. If recall climbs across the bands, you have a dose response pattern that is much harder to explain away as selection, and it is the most persuasive thing a study of this kind can produce. If recall is flat across bands while the exposed group still beats the unexposed one, the difference is probably telling you about who those people are.

Where this stops

Everything above rests on an observational split, and the biggest threat is the one that never shows up in the confidence interval.

Selection is the first problem. People in the hall with the sponsor's branding are there because that hall holds their sector, and their sector is the sponsor's sector. Some part of any lift you measure is that overlap. Banding by exposure intensity helps, and adding a pre-registration awareness question to your registration form helps more, because it gives you a baseline measured before anybody arrived. Neither turns it into an experiment.

Non-response is the second. The people who answer a post-show survey are more engaged than the people who ignore it, and engaged attendees notice sponsors. That inflates both groups, and it inflates them by an unknown and probably unequal amount.

The third is timing. Recall decays, so a survey fielded on the Monday after close and one fielded three weeks later are measuring different things. Pick a day, use the same day every edition, and write it in the method note beside the interval, the same way you would write down the error band on an impression estimate.

None of that argues against running the study. It argues for a report that carries the interval, the decoy rate, the response rate and the field date next to the headline, and for agreeing at contract stage that a recall lift is the objective, which is where the measurement schedule and the report format both start. It is the same exhibitor analytics discipline applied to a survey instead of a scan file.

Before commissioning anything, take your last post-show survey and count how many completes you got and what share of respondents had a badge scan in each hall. That tells you the group sizes available to you, and the margin arithmetic above tells you in ten minutes whether a lift study on your show can answer the question the sponsor is asking.