We analysed 90,651 exhibitor listings. Nine in ten companies never exhibit twice

In a corpus of 90,651 exhibitor listings drawn from 734 trade show editions, 91.2% of the 77,573 distinct companies appear at exactly one show. Take the 108 editions with 50 or more exhibitors and compare every pair of distinct shows: 82.5% of those pairs share literally zero companies, and the median Jaccard overlap is 0.000.

We expected a trade show circuit — a recognisable population of companies that works a run of events every year. It is not there. Below is the number, the query that produced it, the independent check that made us believe it, and the several ways it could still be wrong.

The one-and-done number

Normalising on the lowercased, trimmed company name, the corpus has 77,573 distinct companies. 9,463 of them (12.2%) appear at two or more editions. But editions are not shows: IMEX Frankfurt 2022, 2023 and 2024 are three rows in events, so a company that shows up every year looks like a repeat exhibitor without ever having attended a second event. Collapsing editions into show families by stripping the year out of the event name:

WITH fam AS (
  SELECT id, LOWER(TRIM(REPLACE(REPLACE(name,'2024',''),'2025',''))) AS f
  FROM events
), n AS (
  SELECT LOWER(TRIM(x.company))     AS c,
         COUNT(DISTINCT x.event_id) AS editions,
         COUNT(DISTINCT fam.f)      AS families
  FROM exhibitors x JOIN fam ON fam.id = x.event_id
  GROUP BY 1
)
SELECT COUNT(*)           AS companies,       -- 77573
       SUM(editions >= 2) AS repeat_editions, -- 9463
       SUM(families >= 2) AS repeat_shows     -- 6860
FROM n;

2,603 of the 9,463 apparent repeat exhibitors (27.5%) never attended a second distinct show; they attended the same show in more than one year. Genuine cross-show repeats are 6,860 companies, 8.8% of 77,573. Normalising on registered domain instead — stricter, but only applicable to the 26,216 rows (28.9%) that carry one — gives 23,440 distinct domains of which 1,870 (8.0%) appear at two or more editions. Two different normalisations, same order of magnitude.

The check: 82.5% of show pairs share nothing

Company-level counts are only as good as the name normalisation, so we ran an independent test that does not depend on any single company resolving correctly. Take the 108 editions holding 50 or more rows, build the set of normalised exhibitor names for each, and compute pairwise Jaccard over the 5,735 pairs that are not two editions of the same show.

Jaccard overlap between exhibitor sets, editions with ≥50 rows

Name normalisation errors inflate the number of pairs that appear to share nothing, so treat 82.5% as an upper bound. But the p99 is 0.0105: even at the 99th percentile, two large shows in this corpus have one exhibitor in common per hundred. Whatever is driving that is not a spelling problem, because a spelling problem would have to be nearly universal to move a distribution that far.

The tail is where it gets interesting

The 48 pairs that do overlap heavily are not a gentle continuation of the distribution; they are a different phenomenon. The maximum, at Jaccard 0.937:

Highest-overlap pairs of distinctly-named editions

White Label World Expo and White Label Expo UK share 133 companies out of 136 and 139 — 98% of the smaller list. Those are two separately branded, separately marketed, separately ticketed events with essentially one exhibitor roster. Some of the others are our own deduplication failing (Beauty Istanbul 2024 and BEAUTYISTANBUL 2025 are the same show and our year-stripping heuristic missed the capitalisation change, which is exactly the kind of error we said to expect). Others are real: EUROGUSS and NürnbergMesse's Austrian edition share 447 companies because the same operator runs both. Spring Fair and Autumn Fair share 106 because they are the same UK giftware show twice a year under different names.

So the shape is: a very large population of companies that exhibits once, a negligible middle, and a small number of event pairs that are functionally the same event sold twice. That is a more specific and more useful claim than "exhibitors don't repeat".

Why this is a lower bound, and by how much

The obvious objection is coverage. If we hold 382 editions with data and the world has tens of thousands, then a company exhibiting at two shows will look one-and-done unless we happen to hold both. That is true, it is the largest source of error here, and it biases in a single direction: the true repeat rate is higher than 8.8%.

What we can say is that the effect does not explain the pairwise result. The 108 editions in the Jaccard test are the ones where our coverage is deepest — full published lists, hundreds to thousands of rows each, many of them the flagship show for their industry. If a circuit existed, pairs like Integrated Systems Europe and InfoComm — the two largest pro-AV shows in the world, both fully covered here — should overlap heavily. They share 225 companies out of 1,751 and 855, a Jaccard of 0.094. That is not a coverage artefact.

Our reading is that trade show exhibiting is dominated by regional and single-vertical companies for whom one show is the entire annual plan, and that the multi-show exhibitor is a real but small class. Consistent with that, the companies that do work the circuit are exactly who you would guess:

Most editions attended, exact-string name normalisation

Six companies out of 77,573 appear at ten or more editions. The circuit, such as it is, is enterprise vendors with standing field-marketing budgets, one venue operator that exhibits at other people's shows, and a national trade promotion agency running country pavilions.

Supporting material: what the corpus looks like

The rest of this is context for the numbers above rather than findings in their own right, but it bears on how much you should trust them.

Half the editions publish nothing

Of 734 editions, 382 (52.0%) carry at least one exhibitor row; the other 352 have none. Those are shows where the organiser never published a directory, put it behind a registration wall, rendered it in a way our scrapers could not read, or took it down after the event. Among the 382 that do publish, the mean is 237.3 and the median is 8 — we quote the mean only to note that it is meaningless here.

Exhibitor count per edition, n=382 editions with at least one row

Ten editions hold 28,086 of the 90,651 rows (31.0%); thirty-seven editions — 5% of the corpus — hold 64,734 (71.4%). This is why the Jaccard test is restricted to editions with 50+ rows: below that threshold an edition is a fragment, and a fragment cannot overlap with anything.

Field coverage is decided by the hosting platform

Field population across all 90,651 scraped rows

That 28.9% aggregate hides a bimodal split. Broken out by which ingest source produced the row:

Field coverage by ingest source

MapYourShow gives booth numbers on 99.6% of 22,645 rows and website URLs on none of them. That is not 42 organisers independently choosing to withhold URLs; it is one directory product whose public listing template omits the field. At the per-edition level the same switch shows: of the 108 editions with 50+ rows, 63 have domain coverage of exactly 0% and 26 have coverage above 90%, with 19 anywhere in between.

Separately, we fingerprinted the exhibitor-directory software for 2,778 shows listed on a public event index by resolving CNAMEs and matching homepage markers. We could identify a platform for only 419 of them (15.1%) — the rest were unknown or unreachable, and that 85% is a real limit on this particular claim. But among the 419 we could identify, three vendors account for 398 (95.0%): asp_events 235, MapYourShow 123, a2z 40. If that ratio holds outside our sample, a handful of SaaS companies decide what is knowable about trade show exhibitors, and none of them are in the data business.

Entity resolution is worse than the domain field suggests

Of the 1,870 domains appearing at two or more editions, 636 (34.0%) are written under more than one company name across those editions. The domain field is also not reliably a company identifier: gov.in appears under 17 distinct company names, naver.com under 11, vip.163.com under 9 — those are government portals and consumer email hosts that exhibitors entered where a website was asked for. siemens.com appears under 7 name variants, which is the ordinary case.

Caveats

• The repeat rate is a floor. We hold 382 editions with data. A company exhibiting at two shows we do not both hold is invisible as a repeat. 8.8% is a lower bound; we do not know the ceiling.
• Normalisation is exact-string. Lowercase and trim, nothing else. It does not merge "Siemens AG" with "Siemens Digital Industries", or the same company with and without "GmbH". Every fuzzy scheme we tried also merged genuinely distinct companies, so we chose to undercount. This biases the repeat rate down and the zero-overlap share up.
• Show families are a string replace. Stripping four-digit years from event names is crude, and we showed it failing on Beauty Istanbul above. It splits some shows that should be merged, which inflates the apparent number of distinct shows and therefore the number of pairs.
• Sampling bias toward scrapable platforms. We have six ingest sources. Shows on directory software we have not written a parser for are absent entirely. A show is in this corpus largely because it was technically easy to collect, which correlates with the same platform choices that determine field coverage. The source table above is partly a self-portrait of our scrapers.
• Registration is not attendance. A directory records who booked a booth as of the scrape date. Cancellations, no-shows and late additions are invisible.
• Temporal skew. 445 of 734 editions are 2026 events, many scraped before the show opened and therefore incomplete. Those lists will grow, which will raise both counts and overlaps.

The cleaned corpus behind this post — 637 editions, 5,648 exhibitor records — is searchable and CSV-exportable at exhibitorlens.com; every event page has a free preview.