Exhibitor counting

This page documents how Messe Atlas attributes a single integer to "how many exhibitors were at this Event Edition." Every exhibitor count Atlas publishes — on /events/[slug], in the Briefing, in quarterly reports — links back here.

What we count

For a given Event Edition, Atlas counts the number of distinct exhibiting subsidiaries. That definition has three deliberate choices:

1. Subsidiary, not parent — Henkel Adhesives at one composites show and Henkel Oral Care at IDS Cologne are two different exhibitor records, even though both roll up to Henkel AG. The buyer-relevant unit is the subsidiary, not the holding company.
2. Distinct, not booth-count — if a single subsidiary takes two booths at the same show, that's one exhibitor. Booth count is a separate metric (we may publish it, but it isn't the headline number).
3. Per-edition, not cumulative — the count attaches to one specific year-instance. Year-over-year comparison is the exhibitor churn layer, computed from these per-edition counts.

Sourcing priority

When the same edition's exhibitor list is available from multiple sources, Atlas resolves to one canonical count by source priority:

1. Organizer-published structured list — when the organizer publishes a JSON / CSV / structured HTML exhibitor directory, that's authoritative. Atlas's scrapers parse the structured form directly.
2. Organizer-published prose page — when only an HTML page exists, Atlas's scraper extracts exhibitor names + booth numbers via deterministic parsing or LLM-assisted extraction with manual review for high-priority shows.
3. Press release / aggregated industry coverage — when no organizer publication exists (smaller shows, regional events), Atlas may use trade-press coverage as a starting point with a confidence flag. We don't headline these without verification.
4. Operator-uploaded CSV — for events where the organizer or BD team has direct access, manual upload through the operator UI.

We never publish a count that hasn't been linked to at least one source URL on record.

Deduplication ladder

A scraped exhibitor list typically has 2–5% raw duplicates from name variations. Atlas's entity-resolution pass handles them in a fixed ladder:

1. Domain match — henkel.com is one entity even if the directory lists "Henkel Adhesives Asia Pacific" and "Henkel Adhesive Technologies"
2. LinkedIn URL match — backup canonical when domain is missing
3. Exact-name match after normalization — strip punctuation, legal suffixes (Co/Ltd/Inc/GmbH/SA/SAS), case-fold, collapse whitespace
4. Trigram fuzzy match (sim 0.85+) — borderline pairs that pass automatic merge
5. LLM-assisted disambiguation (sim 0.6–0.85) — Claude reviews each pair with structured output; confidence ≥0.8 merges, lower keeps separate
6. Operator review — anything ambiguous goes to the operator merge queue

Edge cases the count handles

Booth-share / co-located exhibitors. Two formerly separate exhibitors share one booth this year. Both stay in the count; a co-booth reference is stamped on each for analytics.

Acquired companies. Acme exhibited 2023 → BigCo acquired Acme → BigCo exhibits 2025 as itself. Atlas's entity resolution backfills the absorption so churn analytics see continuity.

Pavilions and group stands. A national pavilion (e.g. "Korea Pavilion") usually contains 10–40 sub-exhibitors. Atlas counts each named sub-exhibitor as one. We also stamp pavilion_parent so we can roll up if needed.

Re-brands. If an exhibiting subsidiary is renamed (e.g. Facebook → Meta), Atlas keeps the same subsidiaries.id row and updates the name; participation history follows.

What the count is NOT

To be unambiguous:

• It is not the booth count
• It is not the floor-area count
• It is not the registered-attendee count (that's a separate field on each edition)
• It is not the unique-parent-company count (Henkel + 5 subsidiaries = 6 exhibitor rows, not 1)

When a publication needs one of those instead, Atlas can produce them on request. The default headline number is distinct exhibiting subsidiaries.

When we publish "approximate" instead of an exact count

If our deduplication confidence isn't high (e.g. an event whose list we can only partially reconcile), we publish a banded estimate ("≈800 exhibitors") instead of a point estimate. The methodology page on the specific event surfaces this when relevant.

Cite this

Atlas Methodology, "Exhibitor counting", /methodology/exhibitor-counting.