Cross sell across an event portfolio starts with one account view

A portfolio director asks the commercial team for the number of exhibitors who exhibit on more than one of the group's shows. The answer that comes back is 4 per cent, and everybody in the room knows it is wrong, because three of them can name accounts personally that hold stands on four shows.

The 4 per cent came from matching exhibitor names across show databases. Which found the accounts that happened to be typed identically in two systems and missed everything else. The same manufacturer appears as its UK sales subsidiary on one show, its German parent on another, a brand name on a third, and a distributor holding the stand on a fourth, and no string comparison joins those four without help.

Every cross-sell programme in an event portfolio runs into this before it runs into anything else, and no amount of renewal and cross-sell analytics gets past it. You cannot recommend a second show to an account you cannot see as one account.

The counting problem, before the selling problem

Work it through with numbers, because the shape of the error matters more than its size.

A group runs eight shows. Across the current editions there are 3,100 exhibiting relationships, meaning 3,100 rows of company-holds-stand-on-show. Deduplicate by exact company name string and you get 2,950 distinct names, which implies 3,100 divided by 2,950, or 1.05 shows per exhibitor. Four per cent hold more than one.

Now resolve properly. Normalise legal suffixes, match on name similarity plus domain plus billing address, and roll subsidiaries up to a parent where you can establish the relationship. Say that collapses to 2,400 parent companies. Now the average is 3,100 divided by 2,400, which is 1.29 shows per parent, and the share holding more than one is closer to 22 per cent.

The number did not change because the business changed. It changed because you stopped counting strings and started counting companies. Every downstream figure moves with it: penetration, the size of the cross-sell opportunity, the revenue concentration in your top accounts, and what a lost account actually costs you across the portfolio. So does the retention comparison between single-show and multi-show exhibitors, and whether that gap survives matching is G28's question.

The gap between 1.29 and 2.0 is where people go next, and it is worth being careful. Multiplying 0.71 by 2,400 gives 1,704 additional relationships, which is not headroom. It is an arithmetic artefact of an arbitrary target. Most of your parents have no business being on a second show, because you do not run a second show in their category.

How big is the cross sell opportunity really?

The bounded version takes four steps and each one removes accounts.

Of the 2,400 parents, say 1,850 hold exactly one exhibiting relationship. Those are the population.

Of those 1,850, how many have at least one other show in the portfolio whose category profile genuinely overlaps their product set? Not "we run an industrial show and they make industrial things". A real category match, assessed against the show's exhibitor taxonomy. In most portfolios that removes a third to a half. Say 1,180 remain.

Of those 1,180, how many are large enough to fund a second stand? An exhibitor taking 9 square metres on your smallest show is not buying 40 on your flagship. Filter on current spend, say above 8,000, and you might have 740.

Now apply a conversion rate you can defend. First-year cross-sell conversion on a warm account with a rep call is not 30 per cent. Eight per cent is a number I would put in a plan and be prepared to defend. That gives 59 new relationships. At an average first-edition spend of 14,000, that is 826,000 of new revenue, and at a 45 per cent contribution margin, 372,000 of contribution.

Compare that to the unbounded version, 1,704 relationships at 14,000, which is 23.9 million. One of these two numbers gets a programme funded and then quietly killed in year two. The other one is worth planning against.

How do you resolve the parent company?

Company matching is a different problem from person matching, because the failure mode is different. Nobody mistypes a company name into something unrecognisable. They write a shorter version, a longer version, a divisional version, or the brand instead of the entity. Token overlap after stripping legal suffixes handles that far better than character-level string distance, and the identity resolution machinery that sits under it deserves its own treatment rather than a paragraph here.

The parent relationship is the harder half and it is not a string problem at all. Acme Automation Ltd and Acme Automation GmbH share a name and might be unrelated. Two companies with nothing in common in their names might be under the same owner.

There is a published answer worth knowing about. The Global Legal Entity Identifier Foundation publishes Level 2 relationship data alongside the Legal Entity Identifier, the 20-character code defined by ISO 17442. Level 2 exists specifically to answer who owns whom: each entity with an LEI reports its direct accounting consolidating parent and its ultimate accounting consolidating parent, with the definition anchored in accounting consolidation rather than anyone's opinion of a corporate structure. Coverage is strongest where financial regulation has driven LEI adoption and thinnest among the privately held mid-sized manufacturers who fill most exhibition halls, so this is a source to join against and not a solution on its own.

Which is why I would do this by hand at the top. Take the 200 parents with the highest total portfolio spend, put a person on them for two days, and establish the hierarchy manually with the sales team in the room. Those 200 accounts are probably 45 to 55 per cent of your portfolio revenue, the sales team already knows most of the answers, and the record you produce is correct rather than probable. Let the algorithm handle the tail, where the cost of an occasional wrong merge is small.

Doing the reverse, running an algorithm across the whole file and asking the sales team to review 2,400 proposed merges, is how these projects die.

Two days on 200 accounts sounds like a lot of manual work until you price the alternative. A clerical review queue of 2,400 candidate pairs, at even three minutes a pair, is 120 hours, and it produces decisions made by whoever is free rather than by whoever knows the account. The manual pass also generates something the algorithm cannot: a written note explaining why the German entity and the UK entity are the same commercial relationship, which is the thing a rep needs when the two contacts give contradictory answers about who signs.

Whatever route you take, keep the merges reversible and keep the evidence attached. A wrong parent merge in a cross-sell programme has a specific and embarrassing failure mode, which is a rep phoning an account to pitch a show that a different part of the same group already exhibits at, using the pitch for a first-time exhibitor.

Filling the gaps in what you know about an account

Once the account view exists, most of its cells are empty. You know an exhibitor's stand size, invoice history and service tickets on the one show they take. You know nothing about their category interests, their budget cycle, or whether they exhibit anywhere else. Estimating what they spend with other organisers is G27's problem and it uses the same machinery.

Kamakura, Wedel, de Rosa and Mazzon set out the approach for this in the International Journal of Research in Marketing in 2003. Their mixed data factor analyser combines survey responses from a sample of customers with the transaction data held on everybody, and produces probabilistic predictions of what each customer holds, including holdings with competitors. The mechanism is data augmentation: the survey teaches the model how the observable transaction patterns relate to the unobserved holdings, and the model then imputes those holdings for the customers who never answered a survey.

For an exhibition portfolio the analogue is direct. Survey a sample of exhibitors on which other shows they exhibit at, what their annual exhibition budget is, and which product categories they are pushing this year. Then use the transactional data you hold on everyone, stand size, category, tenure, sponsorship attach, staff registrations, to predict those answers for the accounts who did not respond. The predictions are probabilistic and should be shown as probabilities, because a rep who is told an account probably exhibits at two competitor shows behaves sensibly, and a rep who is told it as a fact and finds out otherwise stops using the system.

What the single account view does not fix

A resolved parent is not always the right unit for selling, and this is the failure I would watch for.

Budget in a large manufacturer usually sits with a division or a country marketing lead, not with the ultimate parent. Roll four divisions of the same group into one account view and you get an accurate picture of total spend and a misleading picture of who decides anything. The rep calls the group marketing director listed on the golden record, who has no budget for trade shows and no interest in the conversation, while the person who actually books stands sits two levels down in a subsidiary you have just hidden.

The version that works keeps both. A parent-level view for measurement, concentration and portfolio reporting, and a buying-unit-level view for selling, with the buying unit recorded explicitly instead of inferred. Distributors and agencies holding stands on behalf of a manufacturer break this further, because the contracting party and the beneficiary are different companies and only one of them is in your billing system.

The second limit is that a cross-sell opportunity is a hypothesis about category fit until somebody tests it. Two shows can look adjacent in a taxonomy and draw completely different buyers, which is a question about audience overlap and belongs with the scoring that picks the next show in G26.

Export the exhibitor list from every show in the portfolio for the last two editions, strip legal suffixes and punctuation, lowercase everything, and count distinct names against total rows. The ratio you get is your current, uncorrected view of portfolio penetration, and the distance between it and the resolved number is the size of the reporting error you are running on today.