Building an exhibitor product taxonomy that a matching engine can use

The matchmaking supplier asks for your category list in week three of the build. Somebody exports the tag list from the exhibitor manual, which has 92 entries, and sends it over. Nobody looks at it closely, because it is the list the show has used for eleven years and it appears on the floor plan and in the printed guide.

Then the first proposals come out and half of them are wrong in a way that nobody can quite name. The exhibitor product taxonomy is where that goes wrong, and it goes wrong before any scoring happens, because product fit is computed over categories and categories are all the engine can see. Two exhibitors selling completely different things sit in the same tag, so the engine treats them as identical. One exhibitor sits under a tag called Other along with 200 others, so the engine has nothing to work with at all.

The taxonomy is the product here. Everything downstream inherits its resolution.

Why is the taxonomy the product?

Because product fit usually carries the largest weight in the score, and product fit is a set comparison over category nodes.

Whatever clever thing the engine does afterwards, it is comparing two lists of category identifiers. If the categories are coarse, every comparison returns coarse answers, and no weighting scheme recovers the detail that was never captured. A taxonomy with 40 nodes over 600 exhibitors puts an average of 15 exhibitors in each node, which means the best possible product fit signal says this buyer wants one of these fifteen. That is not enough to rank a meeting programme, and it is the ceiling, reached only if every exhibitor is classified correctly.

The other reason is that the taxonomy is the shared vocabulary between two populations who never talk to each other. Exhibitors declare what they sell. Buyers declare what they buy. Those declarations only meet if they use the same words, which is I10's problem on the demand side and this post's on the supply side.

What the inherited tag list is actually doing

Take a real shape. A show with 620 exhibitors and 92 tags, where exhibitors select a median of six tags each, so roughly 3,720 tag assignments in total.

Two counts tell you what you have. First, the catch-all. Count exhibitors whose only selections are Other, General, Miscellaneous or the name of the show itself. If 210 of 620 sit there, a third of your floor is invisible to matching and no amount of scoring work touches them.

Second, the distribution. Sort tags by how many exhibitors carry them. On a list that grew by accretion, the top 20 tags typically carry around 70 per cent of the assignments, which would be roughly 2,600 of the 3,720, while the bottom 40 tags carry three exhibitors or fewer each. Both ends are broken in opposite directions. A tag with 180 exhibitors cannot discriminate. A tag with two exhibitors cannot be selected by a buyer who has never heard of it.

The inherited list is also carrying jobs that have nothing to do with matching. It drives floor plan colours, the printed guide index, and the sales team's territory split. Those jobs want a short, stable, marketable list. Matching wants a long, precise, boring one. Trying to serve both with one list is why the list ended up as it is.

Sizing the tree from your own floor

The number of leaves is derivable rather than a matter of taste.

Decide how many exhibitors you want per leaf. Too many and the leaf cannot discriminate. Too few and buyers select a category with nobody in it. Somewhere around eight works for a show of this size, giving a buyer selecting three categories a candidate pool in the mid twenties before any scoring.

Then multiply. 620 exhibitors selecting a median of four leaves each is 2,480 assignments, and at eight exhibitors per leaf that is 310 leaves. Round to 400 to leave room for growth and for the parts of the floor that genuinely need more detail. Above the leaves sit two more levels, roughly 80 mid nodes and 12 top nodes, and the ratios matter less than the fact that every leaf has a parent and every parent has children that make sense together.

The three level structure earns its place through fallback. When a buyer's leaf has four exhibitors in it, the engine climbs to the parent and finds 30, which is the mechanism that keeps a deep tree usable, and choosing which level to match at is I9's subject.

Rebuild the file above onto that tree and the catch-all count is the number to watch. Going from 210 exhibitors in Other to 48 means 162 exhibitors became matchable, which is 26 per cent of the floor, and it is the single largest improvement available to a matching programme that has never had a proper tree.

What a general purpose scheme teaches you

Look at what happens to a classification that has to cover everything. The North American Industry Classification System, revised by the US Census Bureau in 2022, has 1,012 six-digit national industries organised under 20 sectors, and the 2022 revision cut the count from 1,057 by removing distinctions that had stopped being useful, notably between online and physical retail.

Two lessons come out of that. The first is scale: a scheme covering the entire economy of three countries needs only about a thousand terminal categories, so a proposal to build 2,000 leaves for one trade show should be met with suspicion. The second is the revision discipline. NAICS is revised on a five year cycle, with published continuity tables, because a classification that changes continuously cannot support comparison across time and one that never changes stops describing the world.

NAICS is also the wrong tool for the job, which is worth saying plainly, because somebody always proposes it. It classifies what a business is, and matching needs what a business sells at this show. An industrial distributor carrying 40 product lines has one NAICS code and belongs in a dozen of your leaves.

Rules the tree has to obey

Treat it as a controlled vocabulary and use the rules that already exist for those. NISO published Z39.19 in 2005 and reaffirmed it in 2010, covering the construction, format and management of monolingual controlled vocabularies, including lists, taxonomies and thesauri, and it is a free download that will save you a fortnight of arguing.

The parts that matter most on an exhibition floor are these. Each node needs a scope note saying what belongs in it and what does not, because without one, two people classifying the same exhibitor will disagree. Synonyms need to be captured as non-preferred terms pointing at the preferred one, so that an exhibitor typing labelling machinery and one typing labeling equipment land in the same place. Siblings under a parent must be genuinely alternative, and a node should not mix a product type with a material with a market.

One more rule that is local to events. Every leaf needs a name a buyer would recognise without training, because the same tree is what buyers pick from. Internal codes can be as precise as you like. Display labels have about four words to work with.

How do you know the taxonomy is working?

Four counts, run after every edition, none of which need a model.

The catch-all share, which should fall every year. The share of leaves with zero exhibitors, which tells you where the tree is aspirational. The share of leaves with more than 3 per cent of the floor, which on 620 exhibitors means more than 18, and which tells you where the tree is too coarse. And the share of buyer selected categories that have fewer than five exhibitors behind them, which is where you are promising demand you cannot supply.

The fifth measure is the one that closes the loop. For every declined proposal, check whether the buyer and exhibitor shared a leaf, a parent, or nothing. If declined proposals are concentrated in pairs that shared a leaf, that leaf is too broad and should be split. That is the taxonomy's own error report, and it comes free with the meeting data you already hold.

Where a taxonomy stops

A tree assumes an exhibitor can be described by a set of categories, and the largest exhibitors on most floors cannot. A distributor with 4,000 lines will tick 30 leaves honestly and then match against everybody, which degrades the pool for every buyer. Cap the number of selectable leaves, and handle the genuine multi-line distributors as their own top level node with their own rules.

The second limit is time. Product categories move faster than a five year revision cycle in some sectors, and a taxonomy revised annually still lags the floor by a year. Adding leaves mid-cycle is safe when they are children of an existing node, which keeps every prior year's parent-level count comparable. Moving a leaf to a different parent breaks the comparison and needs a mapping table.

The last one is honest and uncomfortable. A better taxonomy raises the ceiling on matching and does nothing on its own, because the ceiling is only reached if the exhibitors are classified correctly against it, and most of them classify themselves in a hurry from a form. That gap is I8's subject, and it is where the effort usually has to go next.

Start by exporting your current tag list with a count of exhibitors against each tag, sorted descending. Read the top ten and the bottom thirty. Then count the exhibitors whose only tags are the catch-all ones. That single number, as a share of the floor, tells you whether your matchmaking has a scoring problem or a vocabulary problem, and it takes twenty minutes to produce.