Building a trade show lead scoring model an organiser can defend to exhibitors

An exhibitor's sales director rings in the second week of December with one question. Why did this lead score 74?

He is not being difficult. He has a file of 900 scored leads, his team can work maybe 200 of them properly before the quarter ends, and he is being asked to trust an ordering produced by somebody else's arithmetic on data he cannot see. If the answer to his question is that the model is proprietary, or that it uses machine learning, he will sort the file by capture time instead and your score will never be opened again.

A trade show lead scoring model an organiser publishes has a harder job than one a company builds for itself. It has to be arguable by a person who did not build it, does not trust you by default, and has a commercial reason to dislike the answer.

What should the organiser score, and what should the exhibitor decide?

Split the two jobs before you write any weights.

The organiser can see things the exhibitor cannot: the registration record behind the badge, the declared job function, the company's own registration history across the portfolio, whether this person registered in September or walked up on the day. That is the organiser's contribution and nobody else can make it.

The exhibitor knows things the organiser never will: which accounts are already customers, which territory has capacity, which product line the sales director needs to move this quarter. That is a threshold decision and it belongs to them.

So the model produces an ordering and the exhibitor draws the line. An organiser who ships a model that also decides which leads are worth calling has taken over a judgement they cannot make and will be wrong in public. That split is the one design rule I would hold to across everything in exhibitor analytics.

There is a legal reason to be careful about the framing as well. Article 4(4) of the GDPR defines profiling as any form of automated processing of personal data used to evaluate certain personal aspects relating to a person, in particular to analyse or predict things like their personal preferences, interests, reliability or behaviour. A score computed from a registrant's job function and behaviour on your floor is that, plainly, and treating it as a neutral piece of reporting because it is only a number for an exhibitor is the wrong instinct. Know your lawful basis, keep the inputs to things the person knowingly provided or did, and be able to describe how the score was produced in a sentence a registrant would understand.

Four inputs, and the arithmetic on one lead

The version I would ship has four inputs, each scored zero to one hundred, combined with stated weights.

Seniority, 0.30. Derived from the job title and function on the registration record, mapped through a published title taxonomy.

Buying role, 0.25. From the registration question about purchasing responsibility, where you ask one. Where you do not, this input does not exist and you should say so rather than substitute a guess.

Category match, 0.25. How closely the registrant's stated interests match the exhibitor's product categories, computed from the same taxonomy that drives the show's own product index.

Qualifier depth, 0.20. How much the stand actually recorded: number of qualifier answers, presence of a next step, a free text note of any length.

Take three leads captured by the same stand.

Lead one, an operations director at a mid-sized processor. Seniority 80, buying role 60, category match 90, qualifier depth 40. The weighted contributions are 24.0, 15.0, 22.5 and 8.0, giving 69.5.

Lead two, a procurement analyst at a large group. Seniority 40, buying role 90, category match 70, qualifier depth 90. Contributions of 12.0, 22.5, 17.5 and 18.0, giving 70.0.

Lead three, a division president who talked for four minutes and said nothing specific. Seniority 100, buying role 80, category match 60, qualifier depth 10. Contributions of 30.0, 20.0, 15.0 and 2.0, giving 67.0.

Three leads inside three points of each other and completely different underneath. Lead two scores highest because the stand did the work. Lead three scores lowest and has the most senior person in it. A sales director who sees 70.0, 69.5 and 67.0 in a column learns nothing. A sales director who sees the four sub-scores beside each total knows within ten seconds which three people these are and which one he is calling first.

Your weights matter less than you think

Somebody will want to argue about 0.30 against 0.35 for seniority. That argument is worth having once and then closing, because the research on linear scoring models says the weights are not where the value sits.

Einhorn and Hogarth, in Organizational Behavior and Human Performance in 1975, examined unit weighting schemes, where every predictor simply gets a weight of one. Their case is that unit weights are not estimated from the data, so they consume no degrees of freedom, carry no estimation error, and cannot reverse the true relative importance of the inputs. Dawes made the broader argument in American Psychologist in 1979, showing across a run of prediction problems that linear models with non-optimal weights, including equal weights, hold up against both optimised regression and expert judgement. The thing that carries the prediction is choosing the right inputs and knowing which direction each one points.

Run the equal weights version on the same three leads. Lead one becomes 80 plus 60 plus 90 plus 40, over four, which is 67.5. Lead two becomes 290 over four, or 72.5. Lead three becomes 250 over four, or 62.5.

Same order, wider spread. The weighted version compressed three genuinely different leads into a three point band, and the equal weights version spread them across ten. That is worth knowing before you tell anyone that a 70 is meaningfully better than a 67.

Then test how fragile the ordering is. Move seniority from 0.30 to 0.33 and take the difference out of qualifier depth, dropping it from 0.20 to 0.17. Lead one goes to 26.4 plus 15.0 plus 22.5 plus 6.8, or 70.7. Lead two goes to 13.2 plus 22.5 plus 17.5 plus 15.3, or 68.5. Lead three goes to 33.0 plus 20.0 plus 15.0 plus 1.7, or 69.7.

The order has completely reversed, from two, one, three to one, three, two, on a weight change of three percentage points that nobody in the room would have defended strongly either way.

The conclusion I draw is about presentation. Ranking inside a narrow band carries no information, and printing it to one decimal place is dishonest precision. Report the score to the nearest five, band the file, and tell the exhibitor plainly that leads within a band are unordered. The stability lives in the extremes, where a 90 and a 35 stay a 90 and a 35 under any weighting anyone will propose. Where those band edges fall is the same argument as setting a letter grade cut point, and it deserves the same procedure rather than a round number.

Render the sub-scores, always

The score that goes into the exhibitor's export should be five columns: the total and the four components, with the weights printed once in the report header.

This is cheap to build and it changes the character of the argument you have with exhibitors. A sales director who thinks seniority is overweighted for his market can now say so with evidence, because he can point at fifteen leads where a high seniority score carried a low category match and none of them converted. That is a better conversation than the one where he tells you the score is wrong and you tell him it is not.

It also surfaces the model's own failures. If category match is returning 50 for every lead at a stand, something in the taxonomy mapping has failed for that exhibitor's product codes, and a flat column in the export announces that immediately, where a slightly wrong total would have hidden it. I have never seen a scoring rollout where that did not happen to at least one category in the first edition.

Print the missing inputs too. A lead where buying role is unknown because the registrant skipped the question should say unknown, and the total should be computed over the inputs you have, with the number of inputs used shown. Substituting a mid-range value for a missing field and calling it a score is how you end up telling an exhibitor that a person is average at something nobody asked them.

How do you check the score against anything?

A score nobody has ever tested against an outcome is a ranking of your own assumptions.

The honest position for most organisers is that they do not have outcome data yet, because the exhibitor holds it and the sales cycle runs past the point where anyone is still interested. That does not stop you doing the cheap checks.

Check the score against the exhibitor's own grade, where a published grading scale has been used. The two are built from different evidence, so they should agree loosely and not exactly. If they correlate almost perfectly, one of them is redundant. If they show no relationship at all, one of them is broken and it is worth finding out which before either goes further. Restrict the comparison to leads that clear the show's qualified lead definition, or you are correlating a score against a pool that is mostly names.

Check the distribution per exhibitor. If one stand's leads score fifteen points below the floor average, that is either a real audience difference or a category match failure, and you want to know which before the report goes out.

Check the score against next edition's registration. Whether a scored lead came back is a weak outcome and it is one you own outright, without asking any exhibitor for anything.

Getting from there to closed revenue is a separate discipline, with the attribution problem and the question of what a show can honestly claim from closed won revenue both needing more care than a scoring model does.

Where this stops

The model scores what the registration record and the scan row contain, and the most predictive thing about a lead is usually the conversation, which appears in neither.

Qualifier depth is a proxy for that conversation and it is a poor one. It measures how much the stand recorded, which tracks how well the stand was briefed at least as much as it tracks how interested the visitor was. A well drilled stand will score its leads higher than an identical stand that never opened the qualifier screen, and the model will read that as better leads rather than better process.

Category match is only as good as the taxonomy underneath it, and product taxonomies at exhibitions are maintained by whoever last updated the show directory. A model built on a taxonomy with sixty categories, twelve of which are used by nobody and three of which mean the same thing, produces category match scores with a lot of decimal places and not much content.

The last limit is the one to state in the report header. Without outcome data, this is a model of what an organiser believes a good lead looks like, expressed transparently enough to be argued with. That is a smaller claim than a predictive model and it is the only one the evidence supports until somebody sends you closed deal data.

Take last edition's scan file for one exhibitor, join it to the registration record, and compute the four sub-scores by hand in a spreadsheet for fifty leads. Then show that sheet to the exhibitor's sales director and ask which of his top ten he would actually call. The gap between his ten and yours is the specification for the next version, and it costs a morning.